Journal of Process Control 24 (2014) 1156–1178

Contents lists available at ScienceDirect

Journal of Process Control

journal homepage: www.elsevier.com/locate/jprocont

A tutorial review of economic model predictive control methods

Matthew Ellis a , Helen Durand a , Panagiotis D. Christofides a,b,∗
a
Department of Chemical and Biomolecular Engineering, University of California, Los Angeles, CA 90095-1592, USA
b
Department of Electrical Engineering, University of California, Los Angeles, CA 90095-1592, USA

a r t i c l e i n f o a b s t r a c t

Article history: An overview of the recent results on economic model predictive control (EMPC) is presented and dis-
Received 3 March 2014 cussed addressing both closed-loop stability and performance for nonlinear systems. A chemical process
Received in revised form 24 March 2014 example is used to provide a demonstration of a few of the various approaches. The paper concludes with
Accepted 24 March 2014
a brief discussion of the current status of EMPC and future research directions to promote and stimulate
Available online 22 April 2014
further research potential in this area.
© 2014 Elsevier Ltd. All rights reserved.
Keywords:
Economic model predictive control
Nonlinear systems
Process control
Process economics
Process optimization

1. Introduction The process performance of a chemical process refers to the

process economics of process operations and encapsulates many
Optimal operation and control of dynamic systems and pro- objectives: profitability, efficiency, variability, capacity, sustaina-
cesses has been a subject of significant research for many years. bility, etc. As a result of continuously changing process economics
Important early results on optimal control of dynamic systems (e.g., variable feedstock, changing energy prices, etc.), process oper-
include optimal control based on the Hamilton–Jacobi–Bellman ation objectives and strategies need to be frequently updated to
equation [16], Pontryagin’s maximum principle [135], and the lin- account for these changes. Traditionally, economic optimization
ear quadratic regulator [84]. Within the context of the chemical and control of chemical processes has been addressed in a multi-
process industries, room for improvement in process operations layer hierarchical architecture (e.g., [106]) which is depicted in
will always exist given that it is unlikely for any process to oper- Fig. 1. In the upper-layer called real-time optimization (RTO), a
ate at the true or theoretically global optimal operating conditions metric usually defining the operating profit or operating cost is
for any substantial length of time. One methodology for improving optimized with respect to up-to-date, steady-state process mod-
process performance is to employ the solution of optimal con- els to compute optimal process set-points (or steady-states). The
trol problems (OCPs) on-line. In other words, control actions for set-points are used by the lower-layer feedback process control
the manipulated inputs of a process are computed by formulat- systems (i.e., supervisory control and regulatory control layers) to
ing and solving a dynamic optimization problem on-line that takes steer the process to operate at these set-points using the manipu-
advantage of a dynamic process model while accounting for pro- lated inputs to the process (e.g., control valves, heating jackets, etc.).
cess constraints. With the available computing power of modern In addition to the previously stated objective, process control also
computers, solving complex dynamic optimization problems (e.g., must work to reject disturbances and ideally, guide the trajectory
large-scale, nonlinear, and non-convex optimization problems) on- of the process dynamics along an optimal path.
line is becoming an increasingly viable option to use as a control The supervisory control layer of Fig. 1 consists of advanced con-
scheme to improve the steady-state and dynamic performance of trol algorithms that are used to account for process constraints, cou-
process operations. pling of process variables, and processing units. In the supervisory
control layer, model predictive control (MPC) (e.g., [116,109,140]),
a control strategy based on optimal control concepts, has been
widely implemented in the chemical process industry. MPC uses a
∗ Corresponding author at: Department of Chemical and Biomolecular Engineer-
dynamic model of the process in the optimization problem to pre-
ing, University of California, Los Angeles, CA 90095-1592, USA. Tel.: +1 310 794 1015;
fax: +1 310 206-4107. dict the future evolution of the process over a finite-time horizon to
E-mail address: [email protected] (P.D. Christofides). determine the optimal input trajectory with respect to a specified

http://dx.doi.org/10.1016/j.jprocont.2014.03.010
0959-1524/© 2014 Elsevier Ltd. All rights reserved.
 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178 1157

performance index. Furthermore, MPC can account for the pro- In an attempt to integrate economic process optimization and
cess constraints and multi-variable interactions in the optimization process control as well as realize the possible process performance
problem. Thus, it has the ability to optimally control constrained improvement achieved by consistently dynamic, transient, or time-
multiple-input multiple-output nonlinear systems. The conven- varying operation (i.e., not forcing the process to operate at a
tional formulations of MPC use a quadratic performance index, pre-specified steady-state), economic MPC (EMPC) has been pro-
which is essentially a measure of the predicted deviation of the posed which incorporates a general cost function or performance
error of the states and inputs from their corresponding steady-state index (i.e., objective function) in its formulation [72,56,141]. The
values, to force the process to the (economically) optimal steady- cost function may be a direct or indirect reflection of the process
state. The regulatory control layer includes mostly single-input economics. However, a by-product of this modification is that EMPC
single-output control loops like proportional-integral-derivative may operate a system in a possibly time-varying fashion to opti-
(PID) control loops that work to implement the computed control mize the process economics (i.e., may not operate the system at a
actions by the supervisory control layer. specified steady-state or target). The rigorous design of EMPC sys-
The overall control architecture of Fig. 1 invokes intuitive tems that operate large-scale processes in a dynamically optimal
time-scale separation arguments between the various layers. For fashion while maintaining stability (safe operation) of the closed-
instance, RTO is executed at a rate of hours-days, while the regu- loop process system is challenging as traditional notions of stability
latory control layer computes control actions for the process at a (e.g., asymptotic stability of a steady-state) may not apply to the
rate of seconds-minutes (e.g., [11,147]). Though this paradigm has closed-loop system under EMPC. It is important to point out that
been successful, we are witnessing the growing need for dynamic the use of OCPs with an economic cost function is not a new con-
market-driven operations which include more efficient and nimble cept. In fact, MPC with an economic cost is not new either (e.g., one
process operation [7,81,150,36]. To enable next-generation opera- such EMPC framework was presented in [72]). However, closed-
tions, novel control methodologies capable of handling dynamic loop stability and performance under EMPC has only recently been
optimization of process operations must be proposed and investi- considered and proved for various EMPC formulations.
gated. In other words, there is a need to develop theory, algorithms, This article attempts to organize the recent theoretical develop-
and implementation strategies to tightly integrate the layers of ments on EMPC. Further explanation of the theory is given where
Fig. 1. The benefits of such work may be transformative to process possible in an attempt to make the theory tractable and accessible
operations and usher in a new era of dynamic (off steady-state) to even a beginning graduate student working in the area of pro-
process operations. cess control. The remainder of the paper is organized as follows.
To this end, it is important to point out that while steady- In the next section, the preliminaries are presented which include
state operation is typically adopted in chemical process industries, the notation used throughout this work, the class of nonlinear pro-
steady-state operation may not necessarily be the economically cess systems considered, as well as a more thorough description of
best operation strategy. The chemical process control literature real-time optimization and model predictive control. The subsec-
is rich with both experimental and simulated chemical processes tions on RTO and MPC are not meant to be comprehensive, but
that demonstrate performance improvement with dynamic pro- rather, are presented to provide some historical background on
cess operation (see [41,94,13,151,149,158,159,131,133,132,153, the challenges addressed in this area. The third section examines
23,97,126,24,105,152], and the numerous references therein for closed-loop stability under EMPC and outlines the various types of
results in this direction). In particular, periodic operation of chemi- constraints and modifications to the objective function that have
cal reactors has been perhaps the most commonly studied example been presented to guarantee some notion of closed-loop stability.
(e.g., [151]). Periodic control strategies have also been developed The fourth section discusses closed-loop performance under EMPC.
for several applications (for instance, [97,126,23,149,133]). Several Various EMPC formulations are subsequently applied to a chemi-
techniques have been proposed to help identify systems where cal process example in the fifth section. An overall discussion and
performance improvement is achieved through periodic operation analysis is provided in the sixth section which attempts to provide
which mostly include frequency response techniques and the appli- our perspective on the current status of EMPC. Lastly, the review
cation of the maximum principle [41,9,21,8,66,158]. concludes with a discussion of future research directions.

2. Preliminaries

2.1. Notation

The operator | · | is used to denote the Euclidean norm of a vec-

tor, while the operator | · |2Q is used to denote a square of a weighted
Euclidean norm of a vector where Q is a positive definite matrix
(i.e., |x|2Q = xT Qx). The symbol S() denotes the family of piece-
wise constant functions with period . A continuous function ˛ : [0,
a) → [0, ∞) belongs to class K if it is strictly increasing and satisfies
˛(0) = 0 and belongs to class K∞ if a =∞ and ˛ is radially unbounded.
A continuous, scalar-valued function, ˇ : Rnx → R is positive defi-
nite with respect to xs if ˇ(xs ) = 0 and ˇ(x) > 0 for all x ∈ Rnx \ {xs }.
The symbol ˝ denotes a level set of a scalar function V(·) (i.e.,
˝ = {x ∈ Rnx |V (x) ≤ }). The set operators ⊕ and  denote the
following set operations:

A ⊕ B = {c = a + b|a ∈ A, b ∈ B}
A  B = {c|{c} ⊕ B ⊆ A}

Fig. 1. The traditional paradigm employed in the chemical process industries for or in other words, A ⊕ B is a set with elements constructed from
process optimization and control. the addition of any element of the set A with any element of the
1158 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178

set B and A  B is a set where the addition of any element of the set recursively solving the process model. If the vector field f(·) is a
A  B with any element of the set B forms a set that is a subset of continuously differentiable function of its arguments, the existence
or is equal to the set A. and uniqueness of this trajectory is guaranteed for all times when
this trajectory is proved to remain within a compact set (e.g., [87]).
2.2. Classes of process systems For the process systems of interest, a continuous function of the
form le : Rnx × Rnu → R is used as a measure of the instantaneous
Throughout this tutorial review, unless otherwise noted, the process operating cost (or profit). As the function le (x, u) is a direct
class of process systems typically encountered within the chem- or indirect reflection of the (instantaneous) process economics it is
ical process industries is considered which are continuous-time typically referred to as the economic cost function, economic cost
systems. Owing to the complex reaction mechanisms and ther- functional, or economic stage cost (here, le (·) will be referred to as
modynamic relationships that govern the underlying physics the economic cost function in subsequent sections). A wide range
of chemical processes, most process systems are inherently of economic costs have been considered such as the net instan-
nonlinear. Mathematically stated, the class of continuous-time, taneous operating profit (i.e., the instantaneous profit minus the
time-invariant nonlinear systems described by the following state- instantaneous cost) as well as more traditional chemical engineer-
space form is considered: ing performance metrics like production rates of desired products,
desired product selectivity, and product yield. Given the general-
ẋ(t) = f (x(t), u(t), w(t)) (1) ity of the classes of the systems encompassed by Eqs. (1) and (2),
further assumptions are placed on the class of systems and are
where x ∈ X ⊆ Rnx Rnu
is the state vector, u ∈ U ⊂ is the manipu-
stated in the subsequent sections as the topics that require these
lated input vector, w ∈ W ⊂ Rnw is the disturbance vector and the
assumptions are introduced.
notation ẋ denotes the time derivative of the state. The set X denotes
The (economically) optimal steady-state is defined to be the
the set of admissible states. The input vector is bounded in the set
minimizer of the following optimization problem:
of the available control energy U where U = {u ∈ Rnu |umin,i ≤ ui ≤
umax,i , i = 1, 2, . . ., nu }. The disturbance vector includes unknown minimize le (xs , us ) (4a)
external forcing of the system, modeling errors, and other forms of xs ,us

uncertainty and is bounded in the following set: W = {w ∈ Rnu ||w| ≤ subject to f (xs , us , 0) = 0 (4b)
,  > 0}.
In addition to continuous-time nonlinear systems, other models g(xs , us ) ≤ 0 (4c)
have been considered for the design of EMPC systems. Specifically, ge (xs , us ) ≤ 0 (4d)
many EMPC schemes have been developed for systems described
by a discrete-time nonlinear model, possibly obtained from the dis- where g : Rnx × Rnu → Rnp denotes the process constraints which
cretization of a nonlinear continuous-time model of the form of may include input and state constraints as well as mixed input
Eq. (1). The discrete-time system analogous to the continuous-time and state constraints and ge : Rnx × Rnu → Rne denotes economic
system of Eq. (1) is given by the nonlinear time-invariant difference constraints like constraints to achieve desired production rates
equation: to meet customer demand, product specifications and quality,
and feedstock availability to name a few. The optimal solution
x(k + 1) = fd (x(k), u(k), w(k)) (2) of Eq. (4) is denoted xs∗ and u∗s . Without loss of generality, the
with x ∈ X ⊆ Rnx , u ∈ U ⊂ Rnu , and w ∈ W ⊂ Rnw where k is used optimal steady-state is assumed to be unique and the origin of
to denote the current time step and the notation fd (·) is used Eq. (1) (i.e., f (xs∗ , u∗s , 0) = f (0, 0, 0) = 0) and similarly, fd (xs∗ , u∗s , 0) =
to distinguish the discrete-time nonlinear state-transition map fd (0, 0, 0) = 0 for the system of Eq. (2).
and the continuous-time nonlinear vector field denoted by f(·). In Remark 1. The state x is assumed to be in the set X ⊆ Rnx (i.e., X
other cases, linear systems are considered. A linear process model may be a subset of or equal to Rnx ). The case that X is the entire Rnx
may arise from the linearization of Eq. (1) around an operating corresponds to the case when no state constraints are considered.
steady-state or when a linear model can provide sufficient accuracy However, the input u is assumed to belong to a set U ⊂ Rnu . The
describing the evolution of the process system. The continuous- assumption that U is only a subset of Rnu is because of the physical
time linear (time-invariant) model is given by limitations of control actuators. Lastly, the disturbance is assumed
to be bounded in a subset of Rnw (i.e., W ⊂ Rnw ) owing to the fact
ẋ(t) = Ax(t) + Bu(t) (3)
that closed-loop stability under a particular control structure in
where A and B are nx × nx and nx × nu matrices, respectively. the presence of disturbances is typically proved for a sufficiently
A state measurement of the process state is assumed to be avail- small bounded disturbance. In general, it is difficult to prove closed-
able at synchronous time instants given by the sequence { k≥0 } loop stability of the closed-loop system of Eq. (1) in the presence of
where  k =  0 + k and  is the sampling period. The kth time possibly unbounded disturbances.
step of the discrete-time model (Eq. (2)) corresponds to the samp-
ling time instance  k of the continuous-time model. To distinguish 2.3. Real-time optimization
between continuous time and the discrete sampling time, the nota-
tion t is used for continuous time and the symbols  k and k are used The traditional method for optimization of chemical processes
for the discrete sampling time instances for the continuous-time is real-time optimization (RTO) (e.g., [59,106,148,56,35]). Typically,
model and the discrete-time model, respectively. Output feedback RTO is executed with a much larger sampling period than the super-
and asynchronous sampling are discussed in Sections 7.1 and 7.3. visory control layer (e.g., MPC layer); that is, RTO may be computed
For the remainder of the manuscript, the predictive controllers on the order of hours-days and the supervisory control layer may
described below will take advantage of the (open-loop) solution to be computed on the order of minutes-hours [106,56,150]. Although
the nominal model (w(t) ≡ 0) of Eq. (1) or (2) for a given piecewise RTO is responsible for process optimization as its name suggests,
constant input trajectory u(t). This state trajectory or solution to the it covers more responsibilities than just optimization in industrial
initial value problem of Eq. (1) with w(t) ≡ 0 for a given initial con- applications. These responsibilities can be summarized in a four-
dition and input trajectory is defined as the open-loop predicted step algorithm. First, the RTO system analyzes process data to detect
state trajectory which is denoted as x̃(t) and can be obtained by if the system has reached steady-state. When steady-state has been
 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178 1159

detected, data validation and reconciliation is completed followed steady-state or the economically optimal trajectory computed in
by model parameter estimation and model updating using vari- an upper-layer optimization problem like the optimization prob-
ous techniques to update the steady-state process model. After the lem of Eq. (4) (e.g., RTO or D-RTO). To manage the trade-off between
model has been updated, optimization, of the form of Eq. (4), is com- the speed of response of the closed-loop system and the amount of
pleted. Lastly, a decision maker decides whether to implement the control energy required to generate the response, MPC is typically
new operating conditions (i.e., send the computed steady-state to formulated with a quadratic objective function which penalizes
the process control layer which forces the process operation to the the deviations of the state and inputs from their corresponding
newly computed steady-state). While RTO has become an impor- optimal steady-state values over the prediction horizon. Specifi-
tant information system in chemical process industries, RTO has cally, MPC is given by the following dynamic optimization problem
three main drawbacks. A complete discussion of the issues arising (recall the assumption that the origin of the system of Eq. (1) is the
in the context of RTO is not within the scope of the present paper, economically optimal steady-state):
but rather a brief summary of these issues is provided below.  N
 
Since optimizing over an accurate process model is important minimize |x̃(t)|2Qc + |u(t)|2Rc dt (5a)
u∈S()
for RTO to yield good performance, RTO has traditionally used 0
more complex nonlinear steady-state models than the supervisory ˙
subject to x̃(t) = f (x̃(t), u(t), 0) (5b)
control layer [60,61]. In the lower feedback control layers, linear
models are often used which may be derived from a number of x̃(0) = x(k ) (5c)
techniques like a linearization of a nonlinear first-principles model
around the desired operating steady-state or via model identifi-
g(x̃(t), u(t)) ≤ 0, ∀t ∈ [0, N ) (5d)
cation techniques (e.g., [20]). The discrepancies between the two where the positive definite matrices Qc > 0 and Rc > 0 are tuning
models may result in a computed operating point by the RTO layer matrices that manage the trade-off between the speed of response
that is unreachable by the feedback control layer often leading to and the cost of control action. The state trajectory x̃(t) is the pre-
an offset between the actual operating steady-state and the desired dicted evolution of the state using the nominal dynamic model
operating steady-state. Also, optimization (or re-optimization) is (w(t) ≡ 0) of Eq. (1) under the piecewise constant input profile com-
completed after steady-state of the process is detected. Since the puted by the MPC. The initial conditions on the dynamic model
process is inherently dynamic and possibly under the influence of are given in Eq. (5c) which are obtained at each sampling period
time-varying disturbances, waiting for the process to reach steady- through a measurement of the current state. The constraints of Eq.
state may delay the computation of the new optimal operating (5d) are the process constraints imposed on the computed input
condition. Thus, re-optimization may be completed only infre- profile (e.g., input and state constraints) which are typically point-
quently, thereby adversely affecting the process performance. One wise constraints, so the constraints of Eq. (5d) are usually written
solution to this problem is to solve the optimization problem more as:
frequently (e.g., [148]), but this may lead to stability issues of the
closed-loop system [56]. g(x̃(j ), u(j )) ≤ 0 (6)
At a more fundamental level, many have questioned whether for j = 0, 1, . . ., N. When the prediction horizon N is finite, it is well-
steady-state operation is the best operating strategy owing to known that the MPC scheme of Eq. (5) may not be stabilizing (e.g.,
time-varying process economics and inherent characteristics of [109]). Various constraints and variations to the cost function may
nonlinear process systems [7,81]. As such, researchers have be made to guarantee stability of the closed-loop system when N is
explored using dynamic models instead of steady-state pro- finite (see, for example, [109], and the references therein).
cess models in the optimization step of RTO, and the resulting To address the drawbacks of the two-layer RTO and MPC hier-
system is typically referred to as dynamic RTO (D-RTO) (e.g., archical control structure, much of recent research has focused
[72,83,107,82,161,176,81,167,169,168]. Dynamic RTO has a simi- on a tighter integration of RTO and MPC. Specifically, to han-
lar structure to that of EMPC, in that both optimization problems dle unreachable set-points, an intermediate layer called the
that characterize these systems are typically dynamic optimization (steady-state) target optimization layer may be introduced that
problems that work to minimize an economic objective subject to a converts the optimal steady-state computed in the RTO layer
dynamic process model. The main differences between D-RTO and to a reachable set-point for the feedback control layer (e.g.,
EMPC are D-RTO is not typically used directly for feedback control, [117,22,172,125,137,171,93,160]). This concept is also referred
but rather it is used in the RTO layer of the hierarchical structure of to as two-stage MPC because of its components. Specifically, a
Fig. 1 above with the process control layers (i.e., supervisory control quadratic program (QP) or linear program (LP) is used to convert the
and regulatory control layers). Furthermore, only limited work has unreachable desired steady-state into a reachable target and then,
been done on a theoretical treatment of closed-loop stability with an MPC of the form of Eq. (5) forces the closed-loop state to the
D-RTO. On the other hand, EMPC is typically implemented for feed- reachable target. Target optimization or the first stage of the two-
back control, its dynamic model is typically implicitly or explicitly stage MPC also allows for more frequent optimization since it is
assumed to be consistent with the model of the optimization layer typically executed at the same rate as the MPC. Within this context,
(e.g., RTO) and its formulation is tailored to account for closed-loop most of the research on this topic has focused on MPC with a linear
stability (see below). model (i.e., using the model of Eq. (3) for the constraint of Eq.(5b)).
Another option is to completely integrate economic optimiza-
2.4. Model predictive control tion of process operations and MPC into the same algorithm. Early
research (and still on-going) on this topic has focused on combin-
Model predictive control (MPC), also referred to as receding ing steady-state economic optimization and linear MPC (i.e., MPC
horizon control, is an on-line optimization-based control tech- formulated with the linear model of Eq. (3)) into one optimiza-
nique that optimizes a performance index or cost function over tion problem. Specifically, MPC schemes that integrate steady-state
a prediction (control) horizon by taking advantage of a dynamic optimization use a cost function of the form:
nominal process model (i.e., Eq. (1) with w(t) ≡ 0) while account- 
ing for process constraints (e.g., [62,116,109,140,137,26]). The main
N
 
LMPC/RTO (x(t), u(t)) = |x(t)|2Qc + |u(t)|2Rc dt + le (x(N ), us )
objective of conventional or tracking MPC is to steer the system 0
to and maintain operation thereafter at the economically optimal (7)
1160 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178

which has both a quadratic (tracking) component and an economic operation may be fixed) are often added. The general formulation
cost component in the cost function. These MPC schemes have the is given by:
following general formulation:
ge (x̃(t), u(t)) ≤ 0 (10)
minimize LMPC/RTO (x̃(t), u(t)) (8a) for all t ∈ [0,  N ). With slight abuse of notation, the constraints
u∈S(),us
of Eq. (10) are not necessarily equivalent to the economics-based
subject to ˙
x̃(t) = f (x̃(t), u(t), 0) (8b) constraints of Eq. (4d), in that they may also incorporate integral,
summation, and average constraints. The constraints of Eq. (10)
x̃(0) = x(k ) (8c)
may play an important role in the solution of the optimization prob-
f (x̃(N ), us ) = 0, us ∈ U (8d) lem of Eq. (9) especially when no upper-layer optimization is used
to account for these constraints and when the optimal operating
g(x̃(t), u(t)) ≤ 0, ∀t ∈ [0, N ) (8e) strategy dictated by Eq. (9) leads to dynamic (transient) operation
(i.e., not steady-state). In either case, the enforcement of these con-
where the decision variables of the optimization problem include
straints in the EMPC is needed to ensure that these constraints are
both the input trajectory over the prediction horizon and the
satisfied over the entire length of process operation.
steady-state input us . The constraint of Eq. (8d) enforces that the
The implementation strategy of the EMPC of Eq. (9) is identical
predicted state trajectory x̃(t) converges to an admissible steady-
to the conventional MPC of Eq. (5). Specifically, EMPC is solved in
state. The remaining constraints and notation are similar to the
a receding horizon fashion. At a sampling instance  k , the EMPC
MPC of Eq. (5). Work in this direction has primarily been appli-
receives a state measurement of the current process state which is
cation driven (e.g., [127,37,173]) with more general frameworks
used to initialize the EMPC. An optimal piecewise input trajectory,
presented in [15,157,3].
according to the optimization problem of Eq. (9), is computed over
the prediction horizon corresponding to the time t ∈ [ k ,  k+N ) in
3. Economic model predictive control schemes
real-time. The optimal input trajectory computed at a given samp-
ling instance is denoted as u* (t| k ). The first control action, denoted
The quadratic cost of conventional MPC (Eq. (5a)) allows for
as u* (0| k ) is sent to the control actuators to be implemented over
tunable closed-loop response. However, it may not be an ade-
the sampling period from  k to  k+1 . At the next sampling period,
quate representation of managing real-time process operation with
the EMPC is re-solved. The resulting input profile computed by the
respect to the process economic performance. A positive deviation
EMPC that is applied to the system of Eq. (1) is denoted as u* (t) and
from the target may represent a profit, while a negative devia-
is given by
tion from the target may represent a loss (or vice versa) [150].
For example, consider an input that supplies heat energy to a u∗ (t) = u∗ (0|k ), for t ∈ [k , k+1 ), k = 0, 1, . . .. (11)
reactor (e.g., a steam jacket). Supplying more steam to the jacket
In the general context (i.e., for the general system of Eq. (1)
than the target is more costly in terms of the energy consumption
or (2)), there are three main issues to consider and address with
of the reactor, while supplying less steam consumes less energy.
respect to the optimal control problem of the EMPC. First is the issue
Owing to this drawback of using a quadratic cost function in the
of feasibility of the optimization problem. Specifically, one must
MPC, the main three drawbacks of RTO, and the calls to unify
carefully consider the conditions that guarantee that the EMPC is
process economic optimization and process control, the idea of
both initially feasible for a given initial condition x( 0 ) and recur-
using the economic cost function le (·) directly in an MPC scheme
sively feasible at each subsequent sampling period. Assuming that
was proposed (e.g., [56,141]). The resulting MPC scheme is called
one can show recursive feasibly, it is important to consider the
economic MPC (EMPC). Since EMPC accounts directly for process
stability or type of stability the closed-loop system will exhibit
economics which is aligned with the core ideas of next-generation
under the EMPC. Recall, no explicit assumption is placed on the
manufacturing (e.g., Smart Manufacturing [31,36], market-driven
economic cost to be positive definite with respect to a steady-state
manufacturing [7], and real-time energy management [150]), its
and thus, the EMPC may dictate a time-varying operating policy.
popularity amongst researchers has significantly increased within
Finally, one should consider the closed-loop performance under the
the last few years.
EMPC. Even though the EMPC optimizes the process economics, it
Broadly, economic model predictive control can be character-
does so over a finite-time prediction or control horizon. Thus, over
ized by the following optimization problem:
long periods of operation, no guarantees, in general, can be made on
 N closed-loop performance under EMPC. For provable results on fea-
minimize le (x̃(t), u(t)) dt (9a) sibility, closed-loop stability, and closed-loop performance under
u∈S() 0
EMPC, additional assumptions must be placed on the closed-loop
subject to ˙
x̃(t) = f (x̃(t), u(t), 0) (9b) system and typically, the addition of stability and/or performance
constraints are added to the formulation of the EMPC. These areas
x̃(0) = x(k ) (9c) are discussed in depth in the subsequent sections.
g(x̃(t), u(t)) ≤ 0, ∀ t ∈ [0, N ) (9d) Several application-oriented formulations of the EMPC of Eq.
(9) have been presented in the literature where an appropriate
where the decision variable to the optimization problem is the cost function and constraints (not of the explicit form discussed
input trajectory over the prediction horizon. The objective function below) have been formulated after an in-depth knowledge of
of Eq. (9a) is the process economic cost function (e.g., operating the application has been gained. The additional elements added
cost) that the EMPC optimizes through dynamic operation of the to the EMPC formulation are tailored for the particular applica-
process. A dynamic model, typically the nominal process model, tion to allow for desirable stability, operation, and performance
is used as a constraint (Eq. (9b)) and is initialized through a state properties. These properties are typically demonstrated and evalu-
measurement obtained at a sampling instance (Eq. (9c)). The con- ated through simulation [56,74,76,75,80,104,1,73]. For theoretical
straint of Eq. (9d) represents process constraints (e.g., input and works that consider EMPC of the form of Eq. (9) (without stability
state constraints) which are implemented as in Eq. (6). In addition constraints), only a limited amount of work has been completed
to the constraints of Eqs. (9b)–(9d), economics-based constraints including [64]. The main advantage of these types of formulations
(e.g., the raw material that may be fed to a process over a period of is that no additional constraints must be added or precomputed. As
 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178 1161

we will see below, the stability constraints are typically obtained certain assumptions on the economic cost function were satisfied.
from a steady-state optimization problem of the form of Eq. (4). This methodology was extended to robust stability of the cyclic
Therefore, the EMPC of Eq. (9) will have less constraints than the steady-state (i.e., input-to-state stability with respect to a bounded
ones discussed below and no steady-state optimization problem is disturbance) [77] where no discount factor was used in the eco-
required to be solved. The main disadvantage of the EMPC of Eq. nomic cost function. In [77], this approach was demonstrated by
(9) is that, at present, stability and performance cannot be guaran- approximating the infinite horizon with a long finite-time hori-
teed in general unless a sufficiently long prediction horizon is used zon which is a typical approach to implement an infinite-horizon
and other controllability assumptions and turnpike conditions are optimal control formulation. In [38], an auxiliary control law was
satisfied [64]. used to formulate the infinite-horizon problem with a finite num-
In the remaining subsections, EMPC formulations with provable ber of decision variables corresponding to a finite-time horizon
closed-loop stability properties are discussed. Closed-loop perfor- and approximating the infinite-horizon tail through the auxiliary
mance under the various EMPC formulations will be discussed in control law. It was shown that the resulting EMPC asymptotically
the subsequent sections. The EMPC formulations are given using stabilizes the economically optimal steady-state when a strong
the continuous-time nominal model (w(t) ≡ 0) of Eq. (1) except for duality assumption is satisfied [38]. Along the same lines as [38],
the EMPC with a terminal constraint. It is straightforward to cast the infinite-horizon EMPC problem was divided into a finite-time
these formulations with the discrete-time nominal model of Eq. (2). horizon and an infinite-horizon tail in [112,113,130]. Of particu-
lar interest to the various applications studied in these works was
Remark 2. It is important to point out that in addition to the
time-varying economic prices. To deal with the infinite-horizon
EMPC schemes discussed below a few EMPC schemes have been
time, an unconstrained infinite-time tail was analytically solved
designed for explicitly handling noise and uncertainty. In [2,67]
for. An economic linear optimal control policy was proposed which
adaptive EMPC schemes were proposed for handling uncertainties
statistically constrained the unconstrained problem and its value
and in [39], an EMPC was presented utilizing stochastic optimiza-
was added to the finite-time horizon EMPC problem as a terminal
tion techniques.
cost [130].

3.1. Infinite-horizon economic model predictive control

3.2. Economic model predictive control with terminal constraints
To address closed-loop stability, one may consider employing
Given the difficulty of solving an infinite-horizon EMPC for
an infinite horizon in the EMPC of Eq. (9). In other words, let N tend
general cost functions of the form le (x, u) and for a general nonlin-
to infinity (similarly, let  N tend to infinity) in the objective function
ear system, a finite-time prediction horizon approach is typically
of Eq. (9a). This is perhaps a more appropriate prediction horizon
adopted. The objective function that the EMPC minimizes is
because chemical processes are continuously operated over long
 N
periods of time (practically infinite time). At least intuitively, one
may be able to guarantee, if a solution is returned, that the state of Le (x(t), u(t)) = le (x(t), u(t)) dt (13)
0
the system is maintained in the set of admissible states X. Then, sta-
bility in the sense of boundedness of the closed-loop state may be where  N = N and N< ∞ is the finite-time prediction horizon. To
guaranteed. Theoretically, one could guarantee that the operating better approximate the infinite-horizon solution and to ensure
policy dictated by the infinite-horizon EMPC is the economically robustness of the control solution to disturbances and instabilities,
optimal one by the principle of optimality. However, it is difficult the finite-horizon EMPC is implemented with a receding horizon;
to solve a general optimization problem with an infinite number of that is, the EMPC optimization problem is solved at every sampling
decision variables. Since optimal control problems such as Eq. (9) instance  k to compute a control action to be applied in a sample-
with N→ ∞ often arise in the context of economics, it is impor- and-hold fashion (i.e., zeroth-order hold) over the sampling period
tant to point out that many ideas for solving various classes of from  k to  k+1 . At the next sampling instance  k+1 , the (finite-
these problems have been proposed, and schemes for obtaining horizon) EMPC is computed by rolling the horizon one sampling
an approximate solution to the optimization problem have been period forward.
devised especially when the open-loop predicted state trajectory Much of the recent theoretical work on EMPC investigates the
displays a turnpike property [28,27,89] which corresponds to the extension of conventional or tracking MPC (Eq. (5)) stabilizing
case in process operations when steady-state operation is likely the elements to EMPC such as adding a terminal constraint and/or
optimal operating strategy (see Section 4 for an illustration of the terminal cost (e.g., see, for instance, [109] for more details on
turnpike property). the use of terminal constraint and cost). Numerous EMPC formu-
Several works on infinite-horizon EMPC have been presented lations and theoretical developments which include a terminal
[169,79,38,77,112,113,130,170]. In [169,170], a few methods were constraint and/or terminal cost have been proposed and studied
given for solving the infinite-horizon EMPC with an economic cost [143,141,58,4,38,79,95,6,119,42,96,142,5,57,63,73,120–122,12,
that maximizes a discounted profit function. In other words, the 166,174]. This class of EMPC schemes has the following general
EMPC is formulated with an objective function: formulation which is given with a discrete-time model as most of
 ∞ the work on this type of EMPC has been done for discrete-time
Le (x(t), u(t)) = − e−t le (x(t), u(t)) dt (12) systems:
0

N−1
where  > 0 is the discount factor used to account for the present minimize le (x̃(j), u(j)) + Vf (x̃(N)) (14a)
value of money. Specifically, a time transformation was introduced u(0),u(1),···,u(N−1)
j=0
to convert the infinite-time interval to a finite-time interval. The
time transformation introduces a singularity which is handled subject to x̃(j + 1) = fd (x̃(j), u(j), 0) (14b)
by imposing a boundary condition at the final time. An adap-
x̃(0) = x(k) (14c)
tive temporal discretization scheme was then employed to solve
the optimization problem. In [79], a discount factor similar to x̃(N) ∈ Xf (14d)
Eq. (12) was used in the economic cost. Nominal stability of
the economically-optimal cyclic steady-state was proved when (x̃(j), u(j)) ∈ Z, ∀ j ∈ I0:N−1 (14e)
1162 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178

where Z ⊆ X × U is a compact, time-invariant set that includes the

process constraints like input and state constraints and I0:N−1 is
the set of integers ranging from 0 to N − 1. At a sampling instance k
corresponding to the time  k in continuous-time, the EMPC of Eq.
(14) receives a measurement of the current state (Eq. (14c)) and
optimizes the economic cost of Eq. (14a) with respect to the pro-
cess dynamics (Eq. (14b)) and process constraints (Eq. (14e)). For
stability and performance (the latter will be discussed in Section 4
below), a terminal constraint is added (Eq. (14d)). If the terminal
constraint is a point-wise constraint x̃(N ) ∈ Xf = {xs∗ }, the terminal
cost, denoted as Vf (x̃(N )), is often dropped as it is not required for
stability and performance guarantees (refer to Section 4 for details
of the latter point). When a terminal region constraint is used, that
is, Xf is some compact set containing xs∗ in its interior, the terminal
cost is often used.
Fig. 2. A state-space illustration of the state trajectory under the EMPC of Eq. (14)
With respect to provable closed-loop stability under the EMPC of
with a point-wise terminal constraint over several sampling periods. The solid line
Eq. (14), an assumption must be placed on the controllability or sta- is the closed-loop state trajectory and the dotted line is the open-loop predicted
bilizability properties of the system of Eq. (2) (or similarly, Eq. (1)). trajectory x̃(t) computed at each sampling period.
Before the assumption can be stated, a few definitions are required.
First, a feasible input solution and the optimal input solution to the x* asymptotically stable for the closed-loop system of Eq. (2) [6,4].
EMPC of Eq. (14) at time step k are denoted as u(0|k), u(1|k), . . ., One method is to make additional assumptions regarding the
u(N − 1|k) and u* (0|k), u* (1|k), . . ., u* (N − 1|k), respectively. The set nonlinear system which extends the notion of dissipativity [6]
of admissible initial states and inputs is the set (first presented for continuous-time systems [165] and extended
ZN = {(x(0), u(0), u(1), . . ., u(N − 1))|x̃(j + 1) = fd (x̃(j), u(j), 0), x̃(0) to discrete-time systems [25]).

= x(0), x̃(N) ∈ Xf , (x̃(j), u(j)) ∈ Z, ∀ j ∈ I0:N−1 } (15) Definition 1. [Dissipativity [25,6]] A closed-loop system is dissi-
pative with respect to a supply rate s : X × U → R if there exists a
where u(j) = u(j|0). The set ZN clearly depends on the prediction function  : X → R such that
horizon length for both a point-wise terminal constraint and a ter-
minal region constraint. The set of admissible initial states, denoted (fd (x, u)) − (x) ≤ s(x, u) (17)
as XN , is the projection of ZN onto X. It is important to note that it for all (x, u) ∈ Z ⊆ X × U. If there exists a positive definite function
is difficult to explicitly characterize the sets ZN and XN in general. ˇ : X → R≥0 such that
The following assumption is placed on the type of discrete systems
considered which bounds the amount of control energy required (fd (x, u)) − (x) ≤ −ˇ(x) + s(x, u) (18)
to force an initial state in XN to xs∗ . The assumption of weak con- then the system is strictly dissipative.
trollability ensures a non-empty feasible set for a sufficiently long
prediction horizon. Assumption 2. [Dissipativity of the Closed-loop System under
EMPC [6]] The closed-loop system of Eq. (2) under the EMPC of
Assumption 1. [Weak controllability] For the system of Eq. (2),
Eq. (14) is strictly dissipative with a supply rate given by:
there exists a feasible input trajectory u(0), u(1), . . ., u(N − 1) for
each x(0) ∈ XN and there exists a K∞ function (·) such that s(x, u) = le (x, u) − le (xs∗ , u∗s ). (19)


N−1 When Assumptions 1 and 2 are satisfied for the closed-loop
|u(j) − u∗s | ≤ (|x(0) − xs∗ |). (16) system under the EMPC of Eq. (14), the steady-state xs∗ is asymp-
j=0 totically stable for any initial condition x(0) ∈ XN . A stronger
assumption than dissipativity that has been used to derive a Lya-
Recursive feasibility of the EMPC of Eq. (14) is guaranteed for
punov function for the closed-loop system under the EMPC of Eq.
the nominally operated system (Eq. (2) with w(k) ≡ 0 and when
(14) is strong duality (i.e., strong duality implies dissipativity [6]).
Assumption 1 is satisfied) for any initial state x(0) ∈ XN . The closed-
loop state trajectory under the EMPC of Eq. (14) will remain Assumption 3. [Strong Duality of the Steady-State Problem [38]]
bounded under nominal operation (w(t) ≡ 0) if the economic cost There exists a s so that (xs∗ , u∗s ) is the unique minimizer of
le (·) and the state transition mapping fd (·) are continuous on Z T
minimize le (x, u) + [x − fd (x, u)] s
(recall that Z is a compact set), xs∗ is contained in the interior of x,u (20)
XN , and Assumption 1 holds. In other words, x ∈ XN for all k ≥ 0 subject to (x, u) ∈ Z
when x(0) ∈ XN (XN is a forward invariant set). This form of sta-
bility is much different than the forms of stability typically shown and there exists a function ˇ ˆ of class K∞ such that the rotated
for the closed-loop system under conventional MPC (e.g., nominal economic cost L(x, u) satisfies:
asymptotic stability when applying the discrete control sequence ˆ
L(x, u)≥ˇ(|x − xs∗ |) (21)
to the discrete-time system of Eq. (2) or practical stability when
applying the discrete control sequence to the continuous-time sys- where the rotated cost is defined as
tem of Eq. (1)). In other words, the EMPC of Eq. (14) will lead to
L(x, u) := le (x, u) + [x − fd (x, u)] s − le (xs∗ , u∗s ).
T
(22)
dynamic, transient, or time-varying operation in general which is
depicted in Fig. 2. This type of stability property has been demon- It was shown in [38] that the rotated cost of Eq. (22) is a Lya-
strated in numerous applications to be an important property of punov function for the closed-loop system under the EMPC of Eq.
EMPC leading to closed-loop economic performance improvement (14) formulated with a point-wise terminal constraint when the
over traditional control methodologies (e.g., tracking MPC). closed-loop system satisfies Assumptions 1 and 3.
Still, it is important to understand under what conditions the Within the context of economics-based constraints, one class
EMPC of Eq. (14) will render the economically optimal steady-state of constraints that are of interest within the context of EMPC are
 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178 1163

average constraints. Reflecting back on the traditional paradigm for Other works on EMPC of the form of Eq. (14) address vari-
economic optimization of chemical processes with RTO, the com- ous issues. In [143,141], the use of unreachable set-points in the
puted operating conditions satisfy the economic constraints (Eq. cost function of a tracking MPC (i.e., with a quadratic cost) was
(4d)) asymptotically since the closed-loop system is asymptotically discussed, and a demonstration of the approach was provided.
forced to the operating conditions. Within the context of EMPC The demonstration showed better closed-loop performance of a
which may not force the system to operate at the economically tracking MPC formulated with unreachable set-points compared
optimal steady-state, it may be important to enforce economics- to using a tracking MPC formulated with reachable targets gen-
based constraints directly in the EMPC. For example, constrain the erated from (steady-state) target optimization (see above for a
EMPC solution such that the time-averaged raw material amount discussion of target optimization). In [58], an EMPC was formu-
is fixed. One notion of economics-based average constraints is to lated for changing economic criterion using a terminal cost and a
construct constraints that asymptotically satisfy an average, as is point-wise terminal constraint of the form: x̃(N) = x̃(N + 1). This
the case in the traditional operation paradigm. type of terminal constraint essentially forces the open-loop pre-
One method for handling asymptotic average constraints [6] is dicted state trajectory x̃ to converge to an equilibrium manifold
to define an auxiliary variable as follows (the method is summa- instead of a single equilibrium point. EMPC with a terminal region
rized below since it is applied in Section 5; the interested reader is constraint and terminal cost was introduced and analyzed in [4].
referred to [6] for a complete discussion of this method): A Lyapunov stability analysis was given in [79] for EMPC with a
terminal constraint based on optimal cyclic steady-states. In [6],
y(k) = h(x(k), u(k)) (23)
asymptotic average performance, the optimality of steady-state
where h : Z → Rny
is continuous on Z and y contains all the average operation, periodic terminal constraints, and asymptotic average
constraints that should be asymptotically satisfied. The average of constraints were presented and analyzed. In [119], it was shown
y is that the dissipativity property is robust to small changes in the con-
n straint set. In [5], direct methods were employed to formulate the
i=0
y(k) EMPC problem as a large-scale nonlinear program (NLP) and solve
(24)
n+1 it with an interior point nonlinear solver [163,19] with automatic
differentiation [164]. The idea of enforcing a generalized termi-
Owing to the fact that h is continuous on the compact set Z, the
nal constraint in EMPC (i.e., enforce the predicted state trajectory
average as n tends to infinity is finite. However, taking the limit
to converge to an equilibrium manifold) was explored further in
of Eq. (24) as n tends to infinity may be meaningless because it
[57,120]. In [57], a MPC (or EMPC) scheme was proposed and it was
may not be properly defined (i.e., have a unique value). Instead, the
shown that with the proposed MPC (or EMPC) algorithm the control
following is used for the definition of the asymptotic average in [6]:
solution converges to that of an MPC (or EMPC) algorithm with a
⎧ ⎫
⎪ 
tn
⎪ terminal constraint chosen to be the economically optimal steady-
⎪
⎪ ⎪
⎪
⎪
⎨
y(k) ⎪
⎬ state. This idea was further extended with a self-tuning terminal
k=0 cost which may lead to improved closed-loop performance com-
Av[y] = y ∈ R |∃tn → ∞, lim
ny
=y (25)
⎪
⎪ n→∞ tn + 1 ⎪
⎪ pared to a fixed terminal weight [120]. In [122], a Lyapunov stability
⎪
⎪ ⎪
⎪
⎩ ⎭ analysis of asymptotically average constrained EMPC was given and
the necessity of dissipativity for optimal steady-state operation was
discussed. A Lyapunov function was derived for EMPC formulated
which deals with the fact that the asymptotic average might be a
with a periodic terminal constraint in [174]. In [166], a two-layer
set of numbers. Also, the asymptotic average of y is a non-empty
control scheme was proposed that featured an EMPC in the upper
set. The (set) constraint on the asymptotic average of y is denoted
layer and a fast hybrid neighboring-extremal controller in the lower
as Y which is assumed to be a closed and convex set. Furthermore,
layer for nonlinear hybrid systems.
it is assumed that

h(xs , us ) ∈ Y. (26) 3.3. Economic model predictive control with Lyapunov-based

At every sampling period, the following constraint is imposed in constraints
the optimization problem of Eq. (14):
Another method for designing an EMPC with provable sta-

N−1 bility properties is to formulate Lyapunov-based constraints
h(x̃(j), u(j)) ∈ Yk (27) by taking advantage of an explicit stabilizing controller (i.e.,
j=0 the explicit controller is used as an auxiliary controller). The
resulting EMPC is the so-called Lyapunov-based EMPC (LEMPC)
where
[68,29,69,54,70,71,50,49,162,51–53,55,91,90,92]. Before the for-

N−1 mulation of LEMPC is given, the main assumptions are stated. The
Yk = Y00 ⊕ (k + N)Y  h(x̃(j), u∗ (j)) (28) vector field f of the nonlinear system of Eq. (1) is assumed to be a
j=0 locally Lipschitz vector function on Rnx × Rnu × Rnw . Like the EMPC
with a terminal constraint, a notion of controllability and/or stabi-
and Y00 is an arbitrary compact set containing h(xs , us ) in its inte- lizability of the system of Eq. (1) must be imposed. The following
rior. The constraint of Eq. (27) ensures that the asymptotic average assumption is essentially a stabilizability assumption for the sys-
constraint is satisfied for the nominal closed-loop system of Eq. tem of Eq. (1) and is comparable to assuming the (A, B) pair is
(2) under the EMPC of Eq. (14) formulated with a point-wise ter- stabilizable for the linear system of Eq. (3) (i.e., the (A, B) pair is sta-
minal constraint (i.e., Xf = {xs∗ } and Vf (·) ≡0). This was shown in bilizable if all its uncontrollable modes are stable or in other words,
[6]. Asymptotic average constraints were extended to EMPC with the eigenvalues of the uncontrollable modes are in the left-half of
a terminal region and terminal cost in [123]. Furthermore, a gen- the complex plane).
eral method for enforcing transient average constraints by adding
(N + T − 1)ny additional constraints to the EMPC where T is the Assumption 4. [Existence of a Lyapunov-based Controller] There
number of sampling periods that the average constraint must be exists a Lyapunov-based controller k(x) which renders the origin
satisfied was presented in [121]. of the nominal closed-loop system of Eq. (1) under continuous
1164 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178

implementation of k(x) asymptotically stable with u = k(x) ∈ U for (30e) and Eq. (30f) define mode 1 and mode 2 operation of the
all x ∈ D ⊆ Rnx where D is an open neighborhood of the origin. LEMPC, respectively. Under mode 1 operation, the trajectory x̃(t)
may dynamically evolve in a bounded set ˝e ⊂ ˝ . The size of
Using converse theorems [108,99,87], Assumption 4 implies the
˝e depends on the stability properties of the system, the samp-
existence of a continuously differentiable Lyapunov function V(x)
ling period, and the bound on the disturbance and has the property
for the nominal closed-loop system of Eq. (1) under the continuous
such that if a disturbance forces the state outside of ˝e over the
implementation of the controller k(x) that satisfies the following
sampling period, the state will be maintained in ˝ . This may be
inequalities:
mathematically summarized through the following: if x(k ) ∈ ˝e ,
˛1 (|x|) ≤ V (x) ≤ ˛2 (|x|) (29a) then x( k+1 ) ∈ ˝ . Under mode 2 operation, the constraint of Eq.
(30f) enforces that the time-derivative of the Lyapunov function
∂V (x)
f (x, k(x)) ≤ −˛3 (|x|) (29b) under the LEMPC be less than the time-derivative of the Lyapunov
∂x
  function under the Lyapunov-based controller k(x). Through this
 ∂V (x)  constraint, the Lyapunov function under LEMPC is guaranteed to
 
 ∂x  ≤ ˛4 (|x|) (29c)
decrease over the sampling period  k to  k+1 for any state x(tk ) ∈ ˝
(thus, x(tk+1 ) ∈ ˝ ). Under mode 2 operation of the LEMPC, the Lya-
for all x ∈ D where ˛i (·), i = 1, 2, 3, 4 are class K functions. The region punov function is guaranteed to decrease until the state trajectory
˝ ⊆ D is the (estimated) stability region of the closed-loop sys- converges to a small neighborhood of the origin as a result of the
tem under the Lyapunov-based controller and is taken to be a level closed-loop stability properties of the controller k(x). The mode 2
set of the Lyapunov function where the time-derivative of the Lya- constraint is enforced to either steer the state to the set ˝e or to
punov function is negative along the closed-loop state trajectory. enforce convergence to the origin (i.e., xs∗ ). An example illustration
Several control laws that satisfy Assumption 4 have been devel- of the possible evolution under the two modes of operation of the
oped for various classes of nonlinear systems including control laws LEMPC is shown in Fig. 3.
that provide explicit characterization of the region of attraction for Mode 1 is active when x(k ) ∈ ˝e and  k < ts , while mode 2 is
the closed-loop system under the controller k(x) which accounts active when x(k ) ∈ / ˝e or  k ≥ ts . The switching time ts warrants
for input constraints (see, for example, [32,48,88,100,156] and the more explanation. The LEMPC scheme may dictate a dynamic oper-
references therein for results in this direction). Applying the con- ating policy. Therefore, continuous forcing of the system through
troller k(x) in a sample-and-hold fashion with a sufficiently small the control actuators may be required to dictate this type of oper-
sampling period will ensure practical stability of the origin of the ation. Therefore, the switching time may be chosen to manage the
closed-loop system (i.e., k(x) is applied as an emulation controller); trade-off between dynamically optimal operation and excess con-
see, for instance, [128,65,118,85,129] and the references therein trol actuator usage. The two extremes, ts = 0 and ts =∞, correspond
for results and analysis of sampled-data systems. Practical stabil- to the case when it is desirable to enforce convergence to the ori-
ity means convergence to a small neighborhood of the origin for gin and to the case when time-varying operation is desirable for
sufficiently large time. the entire length of operation. Notice that the economic cost does
Much like the design methodology of EMPC with terminal con- not need to be modified to enforce convergence to the origin. On
straint/cost, LEMPC takes advantage of design techniques originally the other hand, the cost function is typically modified to achieve
developed for Lyapunov-based MPC (LMPC) which is an MPC tech- guaranteed convergence to the origin under EMPC with a terminal
nique which uses a quadratic cost (e.g., [114,115,118,33]). Utilizing constraint so that the closed-loop system satisfies Assumption 2
the stability region ˝ under the explicit (auxiliary) controller k(x), (accomplished by adding convex or quadratic terms, e.g., [6]).
LEMPC is a two-mode control strategy and its formulation is given The LEMPC has unique feasibility and stability properties com-
by the following optimization problem: pared to EMPC formulated with a terminal constraint. The set ˝
minimize Le (x̃(t), u(t)) (30a) is a characterizable set in state-space and is an estimate of the
u∈S()

subject to ˙
x̃(t) = f (x̃(t), u(t), 0) (30b)
x̃(0) = x(k ) (30c)
u(t) ∈ U, ∀ t ∈ [0, N ) (30d)
V (x̃(t)) ≤ e , ∀ t ∈ [0, N )
(30e)
if V (x(k )) < e and t < ts

∂V ∂V
f (x(k ), u(k ), 0) ≤ f (x(k ), k(x(k )), 0)
∂x ∂x (30f)
if V (x(k ))≥e or t≥ts
where ts is a switching time of the controller which is discussed
below and the other notation is similar to that of Eq. (9).
In the optimization problem of Eq. (30a), the objective func-
tion (Eq. (30a)) is the integral of the economic cost function (Eq.
(13)) over the prediction horizon. The model of Eq. (30b) is used
to predict the future evolution of the process system over the pre-
diction horizon and is initialized through a state measurement at
the current sampling period (Eq. (30c)). The input constraint of
Eq. (30d) bounds the computed piecewise constant input trajec-
tory to be in the set of available control actions. The remaining
Fig. 3. An illustration of the state trajectory under the two-mode LEMPC of Eq. (30a).
Lyapunov-based constraints distinguish the two modes of opera- The state trajectory under mode 1 operation of the LEMPC is the solid trajectory,
tion of the LEMPC. Namely, the Lyapunov-based constraints of Eq. while the dashed trajectory is under mode 2 operation of the LEMPC.
 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178 1165

stability region and feasible set. For any state x ∈ ˝ , the state is number for which V̇ < 0 (under the controller k(x) which accounts
guaranteed to be maintained in ˝ for all times for sufficiently for the input constraint only) may be used. An illustration of the set
small disturbances and a sufficiently small sampling period (˝ is definitions is provided in Fig. 4. Furthermore, the sets ˝x,u and ˝u
forward invariant). Furthermore, the set ˝ does not depend on in Fig. 4 are computed for the example used in Section 5.
the choice of prediction horizon length. Regarding feasibility of Eq. Other theoretical developments on LEMPC include designing a
(30a), the optimization problem is recursively feasible because the state-estimation-based LEMPC using high-gain observers and mov-
input trajectory obtained from the Lyapunov-based controller is a ing horizon estimation [69,55], formulating an LEMPC scheme for
feasible solution to the optimization problem regardless of whether switched systems [71], utilizing LEMPC or Lyapunov-based design
x ∈ ˝ or x ∈/ ˝ . However, stability in the sense of convergence to concepts to design two-layer control structures featuring EMPC or
˝ (and then, boundedness in ˝ ) cannot be guaranteed for any LEMPC [50,51], designing a composite controller with LEMPC for
x∈/ ˝ because the time-derivative of the Lyapunov function under nonlinear singularly perturbed systems [54], accounting for time-
the Lyapunov-based controller may be positive. The set where con- varying pricing in the economic cost function [49], and integrating
vergence to ˝ under the LEMPC is guaranteed (i.e., the region in preventive control actuator maintenance, process economics, and
state-space where the time-derivative of the Lyapunov function is process control into a unified framework with LEMPC [92]. In all
negative under the controller k(x)) is denoted as ˚u and ˝ ⊆ ˚u . the cases, a stability analysis was provided for the system of Eq. (1)
Thus, ˝ is also an estimate of the feasible set. The detailed anal- with bounded disturbances. Additionally, the LEMPC techniques
ysis of the stability properties of this control scheme can be found were applied to parabolic PDE systems along with model reduction
in [68]. techniques [91,90].
Regarding imposing state constraints within LEMPC, one can
extend the concepts from LMPC (e.g., [115]) for imposing state Remark 3. General methods for constructing Lyapunov func-
constraints in LEMPC. Specifically, define the set ˚u as the set tions for nonlinear systems with constraints (e.g., state and input
in state-space that includes all the states where V̇ < 0 under the constraints) remain an open research topic. The construction of
controller k(x). Consider the case where ˚u ⊆ X. This means that Lyapunov functions for unconstrained nonlinear systems may be
any initial state starting in the region X \ ˚u will satisfy the state accomplished by exploiting the system structure like the use of
constraint. However, the time-derivative of the Lyapunov function quadratic Lyapunov functions for feedback linearizable systems
may be positive and thus, it may not be possible to stabilize the and the use of back-stepping techniques. Some methods exist for
closed-loop system starting from this initial condition. The stabil- the design of Lyapunov functions for nonlinear systems with con-
ity region used in the formulation of the LEMPC for this case is straints which include techniques based on Zubov’s method [46]
˝ = ˝x,u = {x ∈ Rnx | V (x) ≤ x,u } where x,u is the largest num- and based on the sum of squares decomposition [134]. In practice,
ber for which ˝x,u ⊆ ˚u . On the other hand, consider the case where quadratic Lyapunov functions have been widely used and have
X ⊂ ˚u . This case is depicted in Fig. 4. For any initial state starting yielded good estimates of the closed-loop stability regions (e.g.,
outside X, the state constraint will be violated from the outset. Also, [32]). While the resulting estimates do not necessarily capture the
for any initial state in the set X, it is not possible, in general, to guar- entire domain of attraction, it is possible to obtain improved esti-
antee that the set X is forward invariant because there may exist mates of the domain of the attraction by using, for example, a family
a stabilizing state trajectory (i.e., a trajectory where V̇ (x) < 0) that of quadratic Lyapunov functions (e.g., [32,49]).
goes outside of the set X before it enters back into the set to con-
verge to the origin. For this case, the determination must be made
whether the state constraints are hard constraints (i.e., cannot be 4. Closed-loop economic performance under EMPC
violated) or soft constraints (i.e., may be violated for some periods
of time). For the case with hard constraints, define the set ˝ as The economic performance of the closed-loop system under
˝ = ˝x,u = {x ∈ Rnx |V (x) ≤ x,u } where x,u is the largest number EMPC is typically measured with the total economic cost index
for which ˝x,u ⊆ X. For the case where the state constraints may be defined by:
treated as soft constraints (i.e., may be violated over certain periods
of time), one can extend the switching constraints of [115] in the  tf
formulation of the LEMPC, since ˚u cannot be computed in practice. Je := le (x(t), u∗ (t)) dt (31)
Instead, the set ˝u = {x ∈ Rnx |V (x) ≤ u }) where u is the largest t0

Fig. 4. An illustration of the various state-space sets described for enforcing state constraints with LEMPC. The case when X ⊂ ˚u is depicted in this illustration.
1166 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178

or the average economic cost index defined by: (defined in Eq. (31)). Under the two cases, the total economic costs
 tf are 29.0 (N = 5) and 15.8 (N = 100), respectively; the performance
1
J e := le (x(t), u∗ (t)) dt (32) with the horizon N = 5 is 84% worse than with N = 100. For the
tf − t0 t0 given system and economic cost, steady-state operation is likely the
optimal operating strategy. The steady-state to which the closed-
where x(t) is the actual closed-loop state trajectory under the input
loop state under the EMPC with N = 5 converges has an economic
profile u* (t) computed by EMPC (i.e., x(t) is the solution of Eq. (1)
cost of 30.0, which is worse than that of the economically optimal
with the input profile u* (t) and a given realization of the process
steady-state which has an economic cost of 15.0. For this initial
disturbances w(t) over the time t0 to tf ). While the EMPC works
condition, any controller that stabilizes around the economically
to optimize the process economics, closed-loop performance has
optimal steady-state would at least asymptotically outperform the
been the subject of much recent research on EMPC. To the casual
EMPC with a prediction horizon of N = 5 from an economic perspec-
observer, it may seem like applying EMPC to a system will result
tive. Furthermore, if we apply the EMPC to the system with various
in improved closed-loop economic performance over traditional
prediction horizons, performance improvement with increasing
control methodologies (e.g., tracking MPC). Unfortunately, this is
prediction horizon is observed (Table 1).
not the case in general. To better illustrate this, consider the simple
The closed-loop trajectory of the system under EMPC may seem
example below.
unexpected or even undesirable. However, the EMPC with N = 5
Example 1. Consider the scalar system described by is performing exactly as it should according to the dynamic opti-
mization problem of Eq. (35). The reason for this behavior is best
ẋ(t) = x(t) + u(t) (33)
explained by observing the open-loop predicted trajectories of the
where the input is bounded by −10 ≤ u(t) ≤ 10 and the economic system under EMPC; that is, the state trajectory under the input
cost for the system is trajectory computed by the EMPC at one sampling period. The
open-loop predicted trajectories for N = 5 and N = 100 are given in
le (x, u) = 2(u + 10) + (x − 2)2 . (34) Fig. 6. Comparing the total economic cost over the open-loop pre-
dicted trajectory from t = 0 to t = 0.25, the total economic costs are
The optimal steady-state and steady-state input are xs∗ = 3.0 and
5.47 with N = 5 and 32.27 with N = 100. While the actions taken by
u∗s = −3.0 which correspond to a steady-state economic cost of
the EMPC with N = 5 are better near-term (over the time period
le (xs∗ , u∗s ) = 15.0. If one formulates an EMPC using the general form
from t = 0 to t = 0.25) compared to the actions taken by the EMPC
of Eq. (9) for the system of Eq. (33), the resulting EMPC would have
with N = 100, these actions are not optimal over a larger horizon.
the following formulation:
 This type of behavior has been observed in many applications (e.g.,
N
  [64,130]) and was described as myopic behavior in [130] which was
minimize 2(u(t) + 10) + (x̃(t) − 2)2 dt originally a term used in the scheduling literature to describe the
u∈S() 0
˙ (35) solution of a scheduling problem derived from an optimal control
subject to x̃(t) = x̃(t) + u(t), x̃(0) = x(k ),
problem that exhibited similar behavior [98].
−10 ≤ u(t) ≤ 10, ∀ t ∈ [0, N ) Another point to be observed from the open-loop predicted tra-
jectory of Fig. 6b is the three distinct segments of the open-loop
where the notation is consistent with the notation used in Eq. (9).
predicted trajectory. The first segment from t = 0 to approximately
The EMPC is applied to the scalar system with a sampling period of
t = 0.5 is the process transients (i.e., the effect of the initial con-
 = 0.05 and two different prediction horizons (N = 5 and N = 100)
dition). From approximately t = 0.5 to t = 4.5, the state trajectory
are considered for a length of operation of t = 5.0. The system is
converges to a neighborhood of the optimal steady-state. In the
initialized at x(0) = 1.0 and the closed-loop trajectories are given
last segment, the state trajectory is driven away from the optimal
in Fig. 5 over the time period t = 0 to t = 0.5 to better illustrate the
steady-state to achieve an improvement in the economic cost. This
difference in the transient operation between the two cases.
property is referred to as a turnpike property [27,40,110,111] since
From the closed-loop trajectories (Fig. 5), the closed-loop sys-
the state passes through the optimal steady-state until it finally
tem under EMPC with the prediction horizons N = 5 and N = 100
moves away to achieve further economic benefit (like a vehicle get-
responds differently which is reflected in the total economic cost
ting on and then, off a turnpike or highway). The turnpike property
is a common property amongst many optimal control problems and
dynamic optimization problems. Unsurprisingly, this property has
been found to be a useful property in the context of EMPC [64,141].

In Example 1, steady-state operation is likely the optimal oper-

ation strategy for the system and the economic cost (in fact, it
will be shown below that it is). Aligned with current practice,
one may consider adding and tuning quadratic terms to the eco-
nomic cost function or additional stabilizing constraints to the
EMPC in an attempt to achieve stabilization at the economically
optimal steady-state. In this case, one potentially helpful tuning
methodology is to observe and understand the open-loop predicted

Table 1
Total economic cost Je with the prediction horizon N.

N Je N Je

1 61928.67 10 15.94
2 361.49 20 15.85
3 70.45 50 15.85
4 41.27 100 15.85
5 28.98
Fig. 5. Closed-loop trajectories of the system of Eq. (33) under the EMPC of Eq. (35).
 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178 1167

Fig. 6. The open-loop predicted trajectories x̃(t) (solid lines) under the optimal input trajectory computed by the EMPC of Eq. (35) with a prediction horizon of (a) N = 5 and
(b) N = 100 (dashed lines are the optimal steady-state and corresponding input).

trajectory of the system under its optimal input solution when solution ufeas = {u∗ (1|k), u∗ (2|k), . . .u∗ (N − 1|k), u∗s }, where u* (j|k)
tuning the stage cost (i.e., the economic cost plus the additional denotes the jth input along the prediction horizon computed at
terms) or assessing what constraints should be added to the EMPC time step k, is a feasible solution to the optimization problem at
to improve performance. k + 1. If there exists any better input solution with respect to the
From a theoretical perspective, currently two methodologies economic cost, the EMPC would return that input solution. How-
exist for closed-loop performance guarantees under EMPC: (1) to ever, using the feasible solution ufeas , the difference in objective
employ a sufficiently large horizon [64] and (2) the application of a function values of the EMPC over two consecutive time steps can
terminal constraint. The former methodology allows for (approxi- be bounded by:
mate) closed-loop performance guarantees for both transient and
infinite-time operation intervals with respect to the economically 
N−1

N−1
optimal steady-state (see [64] for a complete discussion of these le (x(j|k + 1), u∗ (j|k + 1)) − le (x(j|k), u∗ (j|k))
points). For the latter methodology, the type of closed-loop perfor- j=0 j=0
mance guarantee that may be made depends on the type of terminal
constraint used (see Sections 4.1–4.2). The use of a sufficiently ≤ le (xs∗ , u∗s ) − le (x(k), u∗ (j|k)). (36)
large horizon is clearly evident in the above example (Table 1).
Two methods for constructing a terminal constraint are described Using Eq. (36), it was shown in [6] that for nominal operation the
below, but it is important to emphasize these results on closed-loop asymptotic average performance under EMPC formulated with a
performance only work under nominal operation. Closed-loop per- point-wise terminal constraint is bounded above by the econom-
formance under EMPC in the presence of disturbances is an open ically optimal steady-state where the asymptotic average is given
issue. by (defined in discrete time):

T
l (x(k), u(k))
k=0 e
4.1. Terminal constraint or cost lim sup ≤ le (xs∗ , u∗s ). (37)
T →∞ T +1
One approach for closed-loop performance improvement under
EMPC is to consider how to approximate the economic cost that is Asymptotic average performance (Eq. (37)) under EMPC compared
not covered in the prediction horizon (i.e., over the time interval to the economically optimal steady-state has been extended to
[ k , ∞)). To do this, a point-wise terminal constraint based on the EMPC with a terminal region constraint and terminal cost in [4]
economically optimal steady-state Xf = {xs∗ } may be used. At any and to EMPC with a generalized terminal constraint [57,120].
sampling time k (using discrete time), the EMPC is solved under Another important result in the context of performance of EMPC
the constraint that the predicted state at the end of the horizon with a terminal constraint is that when the closed-loop system
converges to the economically optimal steady-state x̃(N) = xs∗ . If under the EMPC with a point-wise terminal constraint is dissipa-
the input trajectory computed at k was applied in closed-loop over tive with supply function s(x, u) = le (x, u) − le (xs∗ , u∗s ), steady-state
the next N sampling periods, the actual closed-loop state would operation is the economically optimal operating strategy [6] (i.e.,
converge to xs∗ at time step k + N where it could be maintained there- no other type of operation will give better average economic cost).
after under nominal conditions. Therefore, the average closed-loop Dissipativity also comes close to being a necessary condition for
economic performance under EMPC formulated with the econom- economically optimal steady-state operation [122]. Unfortunately,
ically optimal steady-state is guaranteed to be no worse than the for large-scale systems dissipativity is hard to verify in general.
economically optimal steady-state over a sufficiently long oper- Since the strong duality condition (Assumption 3) implies dissi-
ating time because over any N time steps, it is possible to force pativity, it can be shown that steady-state operation of the system
the system to xs∗ where it can be maintained thereafter (thus, of Eq. (33) with the economic cost of Eq. (34) is the economically
essentially canceling out the effect of the transients). Of course, optimal operating strategy.
this does not imply that the EMPC formulated with a point-wise
terminal constraint will compute control actions that force the Example 2. For the example system of Eq. (33) under the EMPC
closed-loop state to the steady-state over time (recall the discus- of Eq. (35), the assumption of strong duality (Assumption 3) can be
sion on stability for EMPC formulated with a terminal constraint analytically verified. The state at the next sampling period x( k+1 )
above). This result on the average closed-loop economic perfor- is
mance under EMPC can be mathematically stated as follows: at
the next sampling instance k + 1, the EMPC is re-solved. The input x(k+1 ) = e x(k ) + (e − 1)u(k ) (38)
1168 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178

Applying the definition of the rotated economic cost (Eq. (22)) reactant is fed to the reactor through a feedstock stream with con-
yields: centration CA0 , flow rate F, and temperature T0 . The CSTR contents
are assumed to have a uniform temperature and composition, and
L(x, u) : = le (x(k ), u(k )) + Ts (x(k ) − x(k+1 )) = 2u(k ) the CSTR is assumed to have a constant liquid hold-up. A jacket
+ 20 + (x(k ) − 2)2 + s (x(k ) + u(k ))(1 − e ) (39) provides/removes heat to/from the reactor at rate Q. Applying first
principles and standard modeling assumptions (e.g., constant fluid
The function L(·) is convex and the multiplier s and unique steady- density and heat capacity, Arrhenius rate dependence of the reac-
state minimizer of L(x, u) are tion rate on temperature, etc.), the following system of ordinary
differential equations (ODEs) is derived that describes the evolution
−2
s = , x∗ (k ) = xs∗ = 3 (40) of the CSTR reactant concentration and temperature:
1 − e
dCA F
The property of Eq. (21) is satisfied since L(x, u)≥15 = l(xs∗ , u∗s ) for = (CA0 − CA ) − k0 e−E/RT CA2 (41a)
any x ∈ R and u ∈ U. Therefore, this satisfies the strong duality of dt VR
the steady-state optimization problem which implies that steady- dT F Hk0 −E/RT 2 Q
operation at (xs∗ , u∗s ) is the optimal operating strategy. = (T0 − T ) − e CA + (41b)
dt VR R Cp R Cp VR
Remark 4. Methods exist for approximating the cost-to-go (the where CA denotes the concentration of A in the reactor, T denotes
economic cost over the time interval [ k+N , ∞)) such as estimating the temperature of the reactor contents, and the remaining nota-
this through an infinite-horizon economic linear optimal control tion definitions and process parameter values are given in Table 2.
problem as in [38,130]. The cost-to-go was added to an EMPC for- The CSTR has two manipulated inputs: the inlet concentration of
mulated without a terminal constraint in [38,130]. The approach A with available control energy u1 = CA0 ∈ [0.5, 7.5] kmol m−3 and
demonstrated improved closed-loop performance over an EMPC the heat rate supplied to the reactor u2 = Q with available control
without adding the approximated cost-to-go. energy Q ∈ [−50.0, 50.0] MJ h−1 .
The process economics are assumed to be adequately described
4.2. Performance constraints based on auxiliary controllers by the production rate of the desired product. Therefore, the control
objective of the CSTR is to maximize the production rate of the
As another way to formulate a terminal constraint for closed- desired product while maintaining safe operation of the process
loop performance guarantees under EMPC, one may consider how (i.e., boundedness of the state). The instantaneous economic cost
to construct a constraint which accounts for the closed-loop per- function to accomplish this objective is:
formance over a finite operating window [70,53]. Namely, one may
use an auxiliary stabilizing controller to compute both its input pro- le (x, u) = k0 e−E/RT CA2 (42)
file and open-loop predicted state trajectory over some operating which describes the operating profit and thus, is maximized in the
window. Then, send the terminal state of the computed state tra- EMPC formulations below. Two traditional strategies to increase
jectory (i.e., the state at the end of the operating window with the the production rate are (1) to increase the temperature of the reac-
auxiliary controller) to the EMPC as a terminal constraint. Utilizing a tor contents and (2) to increase the concentration of A by feeding
shrinking prediction horizon, the EMPC computes the economically more A. In the example, the greatest steady-state reaction rate
optimal path to the terminal state. With this EMPC algorithm, the occurs at CAs = 0.143 kmol m−3 and Ts = 711.1 K corresponding to the
closed-loop economic performance under EMPC is at least as good steady-state inputs CA0s = 7.5 kmol m−3 and Qs = 50.0 MJ h−1 with
as the closed-loop economic performance under the auxiliary stabi- steady-state reaction rate le (xs , us ) = 36.8 kmol m−3 h−1 (the units
lizing controller on both the finite-time and infinite-time intervals on the reaction rate are dropped in the remainder). However, from
[53]. Another idea is to compute the total control energy used by a practical perspective, it may not be desirable to operate at such
the auxiliary stabilizing controller and enforce that the EMPC com- a high temperature and/or operate at a steady-state that uses the
putes an input trajectory that uses no more control energy than maximum available control energy. To address this, consider a con-
the auxiliary controller input profile over the operating window straint on the time-average reactant material of the form:
[70]. This may be particularly important when the economic cost  tavg
function does not penalize the use of control energy. 1
CA0 (t) dt = CA0,avg (43)
tavg 0
5. Evaluation of EMPC using a chemical process example
where tavg is the time over which to enforce the material constraint
and CA0,avg. is an average amount of reactant which is taken to
In this section, various EMPC formulations are demonstrated.
be the median value CA0,avg = 4.0 kmol m−3 . This fixes the optimal
Specifically, applications of the various EMPC schemes to a chem- ∗ =C
inlet concentration steady-state value to CA0 A0,avg . Both cases
ical process example are considered in this section. The specific
where the operating period tavg is chosen to be finite and infinite
chemical process example has been chosen because the under-
are considered below.
standing of its dynamic evolution is tractable to most engineers
familiar with chemical processes. This section is not meant to apply
all available EMPC formulations/algorithms presented in the litera- Table 2
Process parameters of the CSTR.
ture to a chemical process example, but rather, to discuss how one
can design an EMPC of the form of Eq. (9) for a specific application Symbol Description Value
by taking advantage of several of the theoretical developments of F Feedstock flow rate 5.0 m3 h−1
EMPC presented in the literature. T0 Feedstock temperature 300 K
VR Reactor fluid volume 1.0 m3
E Activation energy 5.0 × 104 kJ kmol−1
5.1. CSTR description k0 Pre-exponential rate factor 8.46 × 106 m3 kmol−1 h−1
H Reaction enthalpy change −1.16 × 104 kJ kmol−1
Consider a non-isothermal continuously stirred tank reactor Cp Heat capacity 0.231 kJ kg−1 K−1
(CSTR) where an elementary, exothermic second-order reaction R Density 1000 kg m−3
R Gas constant 8.314 kJ kmol−1 K−1
takes place that converts the reactant A to the desired product B. The
 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178 1169

In the simulations below, the open-source interior point nonlin- 

min{k+N, M−1}

k−1
ear optimization solver Ipopt [163,19] was used to solve the EMPC Muavg − u(j ) − u∗ (i )≥ max{M − N − k, 0}umin .
problems at every sampling instance. To numerically integrate the j=k i=0
ODEs of Eq. (41) forward in time, the explicit Euler method was (45b)
used with an integration time step of 0.001 h. The sampling period
of the EMPC schemes presented below is  = 0.01 h. Together these constraints ensure that the average constraint of
Eq. (44) is satisfied. Specifically, Eq. (45) means that the difference
between the total available input energy (Muavg ) and the total input
energy used from the beginning of the operating period through
5.2. Closed-loop performance under EMPC
the end of the prediction horizon must be equal to or less/greater
than the total input energy if the maximum/minimum allowable
In the first set of simulations, closed-loop economic perfor-
input was applied over the remaining part of the operating period
mance under EMPC is considered where two different EMPC
from  k+N to  M . If the prediction horizon extends over multiple
schemes are formulated and applied to the nominally operated
consecutive operating periods, a combination of the constraints of
CSTR (w(t) ≡ 0). The EMPC is allowed to operate the CSTR in a large
Eq. (44) and (45) can be employed. For example, if the prediction
operating envelope, and no state constraints are imposed. Explicit
horizon extends over two operating periods, the constraint to be
considerations of the practicality of the operating policy dictated
enforced at a sampling instance,  k , becomes:
by the EMPC (e.g., consideration of state constraints) is left to the
subsequent subsection which discusses closed-loop stability under 
M−1

k−1
EMPC (Section 5.3). However, to provide some limit to the operat- u(i ) + u∗ (i ) = Muavg , (46a)
ing range of the CSTR the available control energy for the heat rate i=k i=0
input to the CSTR is restricted to Q ∈ [0.0, 20.0] MJ h−1 . This restric-
tion limits the temperature range over which the EMPC operates 
min{k+N, 2M−1}

the CSTR because the optimal operating strategy is to provide the Muavg − u(j ) ≤ max{2M − N − k, 0}umax , (46b)
upper limit heat rate to the CSTR for all time to make the reaction j=M
rate as large as possible (as discussed above). Therefore, the eco-
nomically optimal steady-state is defined as CAs ∗ = 0.719 kmol m−3 
min{k+N, 2M−1}

∗
and Ts = 481.4 K with corresponding optimal steady-state inputs Muavg − u(j )≥ max{2M − N − k, 0}umin . (46c)
∗ = 4.0 kmol m−3 and Q ∗ = 20.0 MJ h−1 and with a steady-
of CA0s j=M
s
state production rate of le (xs∗ , u∗s ) = 16.4. For simplicity of notation, k is reset (i.e., k = 0 at the sampling period
Recall that the time-averaged amount of material fed to the CSTR  M ) at the beginning of each operating period.
is fixed. One method to ensure that the average constraint of Eq. The following EMPC is applied to the CSTR system:
(43) is satisfied over the entire length of operation is to construct  N
a constraint that ensures that the constraint is satisfied over each
maximize k0 e−E/RT̃ (t) C̃A2 (t)dt (47a)
consecutive operating period tavg . This may be accomplished by u∈S() 0
using a simple inventory balance accounting for the total amount
F
of input energy available over each operating period compared to subject to C̃˙ A (t) = (u1 (t) − C̃A (t)) − k0 e−E/RT̃ (t) C̃A2 (t) (47b)
V
the total amount of input energy already used in the operating
period. The main advantages of enforcing the average constraint F Hk0 −E/RT̃ (t) 2 u2 (t)
T̃˙ (t) = (T0 − T̃ (t)) − e C̃A (t) + (47c)
in this fashion are that (1) only a limited number of constraints are V Cp Cp V
required to be added to the EMPC, and (2) it ensures that the aver-
age constraint is satisfied on both the finite-time and infinite-time C̃A (0) = CA (k ), T̃ (0) = T (k ) (47d)
intervals. The enforcement of the constraint is carried out as fol- u(t) ∈ U, ∀ t ∈ [0, N ) (47e)
lows: if the prediction horizon covers the entire operating period,
then the average constraint can be enforced directly; that is, impose 
M−k−1

k−1

the following constraint in the optimization problem of the EMPC: u1 (j ) + u∗1 (i ) = Mu1,avg (47f)
j=0 i=0


M−1

N
u(i ) = uavg (44) Mu1,avg − u1 (j ) ≤ max{2M − N − k, 0}u1,max (47g)
M
i=0 j=M−k

where  M is the operating period length that the average input con- 
N

straint is imposed (i.e., M =  M / is the number of sampling periods Mu1,avg − u1 (j )≥ max{2M − N − k, 0}u1,min (47h)
in the operating period) and uavg is the average input constraint j=M−k
value. The integral of Eq. (43) has been converted to a sum in Eq.
where the notation used is similar to the previous EMPC formu-
(44) because the input trajectory is piecewise constant. If the pre-
lations. We note that the EMPC of Eq. (47) resets its initial time
diction horizon does not cover the entire operating period, then the
to zero at each sampling period (i.e., the real-time horizon  k to
remaining part of the operating period not covered in the predic-
 k+N corresponds to the prediction horizon from 0 to  N in the
tion horizon must be accounted for in the constraints. Namely, at a
controller). Therefore, the time indices of the constraints of Eq.
sampling period  k ∈ [ 0 ,  M ), the following must be satisfied:
(47f)–(47h) are shifted to account for this point. The closed-loop
simulation results are shown in Figs. 7 and 8 for the initial con-

min{k+N, M−1}

k−1
dition CA (0) = 2.0 kmol m−3 and T(0) = 425.0 K, an operating period
Muavg − u(j ) − u∗ (i ) ≤ max{M − N − k, 0}umax , of 100 sampling periods (i.e., M = 100 and  M = 1.0 h) and predic-
j=k i=0 tion horizon N = 10. The closed-loop state trajectories converge to a
(45a) limit cycle over several periods of operation (Fig. 8). The closed-loop
performance of the system under the EMPC of Eq. (47) is evaluated
1170 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178

Table 3
Closed-loop performance over 50 h of operation under the EMPC with a terminal
constraint and the constraint of Eq. (48) with various prediction horizon lengths.

N JE

10 16.76
20 17.03
30 17.05
40 17.09
50 17.14

a terminal (point-wise) constraint based on the economically opti-

mal steady-state is added to the EMPC. However, the average input
constraint must be carefully constructed. Applying constraints of
the form of Eq. (46) will likely cause the EMPC, formulated with
a terminal constraint, to become infeasible owing to the fact that
the input constraint may become tight near the end of the operat-
ing window. The terminal constraint may no longer be a reachable
steady-state with the remaining control energy and thus, increas-
ing the prediction horizon may not resolve this issue for this type of
average constraint. For instance, initializing the CSTR with the ini-
tial condition CA (0) = 2.0 kmol m−3 and T(0) = 425.0 K, the EMPC of
Fig. 7. Closed-loop trajectories of the CSTR over 10 h of operation under the EMPC
with the operating period average input constraints.
Eq. (47) with the terminal constraint x̃(N ) = xs∗ becomes infeasible
at 0.78 h. One method to resolve this issue is to handle the average
constraint asymptotically [6].
using the average economic cost index of Eq. (32). Over the first
One type of constraint to enforce such that the input aver-
operating period under the EMPC of Eq. (47), the average economic
age constraint is asymptotically satisfied is the constraint of Eq.
cost is 19.95, while the average economic cost over one hour start-
(27) (presented in [6]). Formulating constraints of this form which
ing from the same initial condition and under a constant input of
ensure that the average input constraint is asymptotically satisfied
u∗s is 17.46 (operation under EMPC has 14.25% better performance).
for the CSTR example yields the following constraints:
Over the 50.0 h length of operation, the average economic cost
under the EMPC is 12.68, while under the constant steady-state
input value it is 16.43. The average closed-loop economic perfor-

k+N

k−1
u(j )≥Numin + (k + N)uavg − u∗ (j ) (48a)
mance under EMPC is 22.78% worse than the performance under
j=k j=0
the constant input u∗s . It is important to emphasize that several
chemical process examples under EMPC formulated without the
use of a point-wise terminal constraint and without a terminal cost 
k+N

k−1
u(j ) ≤ Numax + (k + N)uavg − u∗ (j ) (48b)
have demonstrated improved economic closed-loop performance
j=k j=0
over traditional control methods (e.g., [71,51]).
One solution to guarantee performance improvement over
steady-state operation for long term (infinite-time) operation is
to add a terminal constraint based on the economically optimal
steady-state or the open-loop predicted trajectory under an auxil-
iary controller which was described in Section 4 above. Therefore,

Fig. 9. The average of CA0 computed by the EMPC with a terminal constraint, a
Fig. 8. The closed-loop state trajectory in state-space of the CSTR over 10 h of oper- prediction horizon length of N = 20 and the asymptotic average input constraints
ation under the EMPC with the operating period average input constraints. (solid trajectory). The desired average, u1,avg is the dashed line.
 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178 1171

Fig. 10. The closed-loop trajectories of the CSTR over 50 h of operation under the Fig. 12. The closed-loop trajectories of the CSTR over 50 h of operation under the
EMPC with a terminal constraint, a prediction horizon length of N = 20 and the EMPC with a terminal constraint, a prediction horizon length of N = 50 and the
asymptotic average input constraints. asymptotic average input constraints.

which replace the average constraints of Eq. (47f)–(47h) in the constraint operates the CSTR in a much smaller operating range
EMPC of Eq. (47). Also, the terminal constraint x̃(N ) = xs∗ is added compared to the CSTR under the EMPC of Eq. (47) (Fig. 7). Interest-
to the EMPC. The resulting EMPC is applied to the CSTR. The ingly, however, operation over a much larger operating region as
closed-loop performance is given in Table 3 for the initial condition is the case under the EMPC of Eq. (47) (with N = 10) does not yield
CA (0) = 2.0 kmol m−3 and T(0) = 425.0 K and for several prediction better closed-loop economic performance compared to operating
horizon lengths. Overall, a similar trend in performance with in a smaller region as is the case under the EMPC with a terminal
prediction horizon length is observed in Table 3 as the one observed constraint and with N = 10. Specifically, the average reaction rate
in Example 1. over the 50.0 h simulation is 16.76 with the EMPC with a terminal
The average CA0 value with time for the input trajectory com- constraint and with N = 10. This is 1.93% better than steady-state
puted by the EMPC with a prediction horizon of N = 20 is given operation (constant input u∗s ) and 31.07% better than the EMPC
in Fig. 9 which demonstrates that the average constraint on CA0 without the terminal constraint.
is asymptotically satisfied. The closed-loop trajectories under the From the simulations of the CSTR under the EMPC with the
EMPC with a terminal constraint are shown in Figs. 10 and 11 for terminal constraint, the computed CA0 profile is approximately a
N = 20 and Figs. 12 and 13 for N = 50. From the closed-loop trajec- periodic profile. From these simulations, the period of the periodic
tories of the CSTR under the EMPC with N = 20 (Fig. 10), a periodic switching policy dictated by EMPC may be approximated. On aver-
operating policy is dictated by the EMPC; a more complex periodic- age, the EMPC with a terminal constraint switches between CA0,max
like operating policy is observed with N = 50 (Fig. 12). Owing to the and CA0,min approximately every 10 sampling periods and thus,
enforcement of the terminal constraint, the EMPC with a terminal the optimal period of switching is approximated as 20 sampling

Fig. 11. The closed-loop state trajectory in state-space of the CSTR over 50 h of Fig. 13. The closed-loop state trajectory in state-space of the CSTR over 50 h of
operation under the EMPC with a terminal constraint, a prediction horizon length operation under the EMPC with a terminal constraint, a prediction horizon length
of N = 20 and the asymptotic average input constraints. of N = 50 and the asymptotic average input constraints.
1172 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178

periods. Utilizing the computed optimal period consisting of 20 terminal penalty to the economic cost while still having provable
sampling periods, we revisit the EMPC of Eq. (47) which uses an stability guarantees in the presence of bounded disturbances (i.e.,
average input constraint enforced over operating windows of size boundedness of the closed-loop state in the region). The region
M. constraint, ˝ , is characterized with an explicit stabilizing con-
Consider the EMPC of Eq. (47) where the operating period is troller, k(x), and therefore, is an estimate of the region of attraction
now 20 sampling periods (M = 20) and the prediction horizon is for the system under the input constraints. Furthermore, different
N = 20. The EMPC is applied to the CSTR with two initial conditions: Lyapunov-based constraints can be formulated to achieve multiple
T = [2.0 425.0] and xT = [1.0 490.0]. The closed-loop per-
x0,1 objectives which will be discussed and demonstrated below.
0,2
formances under the EMPC of Eq. (47) (with M = 20 and N = 20) for Before the auxiliary explicit stabilizing controller can be
these two initial conditions are 16.95 and 16.94, respectively, while designed for the LEMPC, the control objective for the CSTR is mod-
the closed-loop performances under the EMPC with a terminal con- ified. Since the economic cost does not penalize the use of control
straint, the asymptotic average constraint of Eq. (48), and N = 20 are energy, the optimal operating strategy is to operate at the max-
17.03 and 16.99 for these two initial conditions, respectively. The imum allowable heat rate supplied to the reactor. However, this
differences in performance between the two EMPC schemes are may lead to a large temperature operating range (Fig. 8) which
0.48% for the first initial condition and 0.29% for the second, with may be impractical or undesirable. Therefore, consider a modified
the EMPC with a terminal constraint having the better performance control objective for more practical closed-loop operation of the
in each case. However, recall that this comparison does not amount CSTR under EMPC. The modified control objective is to maximize
to comparing two equivalent scenarios. Since the EMPC under the the reaction rate while feeding a time-averaged fixed amount of
asymptotic average constraint only needs to asymptotically satisfy the reactant A to the process and while forcing and maintaining
the average constraint, it may use more (or less) input energy over operation to/at a pre-specified set-point temperature. Additionally,
the finite-time (transient) interval. In fact, for this example, the the temperature of the reactor contents must be maintained below
EMPC with the terminal constraint and asymptotic average con- the maximum allowable temperature T(t) ≤ Tmax = 470.0 K, which is
straint uses an average of 4.01 kmol m−3 for both initial conditions treated as a hard constraint and thus, X = {x ∈ R2 |x2 ≤ 470.0}. The
which is slightly more than the EMPC with operating period average optimal steady-state, in this case, is obtained from the need to sat-
constraints. This analysis suggests that it may be useful, when using isfy the new control objective and therefore, is not directly derived
an operating period average input constraint, to construct and solve from an optimization problem. Specifically, the heat rate input is
a dynamic optimization problem to determine the optimal period allowed to take values in its full set of available control energy
( M ) to enforce the operating period average input constraint since (Q ∈ [−50.0, 50.0] MJ h−1 ) and the optimal steady-state inputs are
∗ =C −3 and Q ∗ is the average
the closed-loop performance under EMPC with an average con- set as follows: CA0 A0,avg = 4.0 kmol m s
straint enforced over operating periods may be dependent on the available heat rate which is Qs∗ = 0 MJ h−1 . The reasoning for the
choice of operating period length. latter choice is to have an equal amount of positive and negative
control energy. The steady-state in the operating range of interest
5.3. Stability under EMPC corresponding to steady-state input values of CA0∗ = 4.0 kmol m−3

and Qs∗ = 0 MJ h−1 is CAs

∗ = 1.18 kmol m−3 and T ∗ = 440.9 K and is
s
In the previous section, closed-loop performance was studied
open-loop asymptotically stable.
comparing the closed-loop performance of the CSTR under EMPC
A stabilizing state feedback controller k(x) is designed for the
formulated with a terminal constraint and under an EMPC with-
CSTR with respect to the optimal steady-state. The first input CA0
out a terminal constraint. For this particular example, closed-loop
in the stabilizing controller is fixed to the average inlet concen-
economic performance of the EMPC without a terminal constraint
tration to satisfy the average input constraint. The second input
and an average input constraint formulated for successive operat-
Q is designed via feedback linearization with a controller gain of
ing windows and an EMPC formulated with a terminal constraint
 = 1.4 (see [68] for more details regarding the controller design). A
and asymptotic average input constraint yield similar closed- T
quadratic Lyapunov function is considered of the form V (x) = x Px
loop performance (for an appropriately chosen operating period
 M of the first EMPC). Furthermore, both cases demonstrated a
closed-loop economic performance improvement under dynamic
operation compared to steady-state operation. In the previous
study, strictly nominal operation was considered (i.e., operation
under no plant-model mismatch, no disturbances, and no other
uncertainty). Practically speaking, it is never possible to achieve
nominal operation. Furthermore, work in the direction of closed-
loop performance under actual operation (e.g., in the presence
of process noise, external forcing and unmeasured disturbances,
communication disruptions between components of the control
architecture, etc.) remains an open research topic.
In this section, operation under process noise and plant-model
mismatch is considered. As argued in [4,141], the use of a (ter-
minal) region is superior to the point-wise terminal constraint.
Arguing for or against this point for EMPC is not within the scope
of this work, but instead, the discussion proceeds using this
point to motivate the use of an EMPC with a region constraint
instead of a point-wise terminal constraint for operation in the
presence of process noise. Specifically, an LEMPC is chosen to be
formulated and applied to the CSTR model due to several of its
unique properties compared to EMPC with a terminal region and
terminal cost (e.g., the EMPC presented in [4]). With LEMPC, a
region constraint can be constructed without the need to add a Fig. 14. Two closed-loop state trajectories under the LEMPC in state-space.
 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178 1173

where x is the deviation of the states from their corresponding Table 4

Average economic cost over several simulations under the LEMPC, the Lyapunov-
steady-state values and P is the following positive definite matrix:
based controller applied in a sample-and-hold fashion, and the economically
  optimal input u∗s . For the case denoted with a “*”, the system under the constant input
250 5 u∗s settled at a steady-state different from the economically optimal steady-state.
P= . (49)
5 0.2
J E under LEMPC J E under k(x) J E under u∗s
The stability region of the CSTR under the controller k(x) is char- 14.17 14.10 14.09
acterized with a level set of the Lyapunov function where the 14.18 14.11 14.09
time-derivative of the Lyapunov function along the closed-loop 14.17 14.10 14.08
14.17 14.09 14.06
state trajectories is negative and is denoted as ˝u = {x ∈ R2 |V (x) ≤
14.18 14.10 14.10
u } where u = 138. However, X ⊂ ˝u which is shown in Fig. 14 and 14.17 14.09 14.08
thus, we define the set ˝ where  = 84.76 to account for the state 14.18 14.09 14.10
constraint. Bounded Gaussian process noise is added to the CSTR 14.18 14.08 14.08
14.19 14.08 14.10
with a standard deviation of = [0.3 5.0]T and bound  = [1.0 20.0]T .
14.18 14.07 14.07
Specifically, a new noise vector is generated and applied additively 14.18 14.11 14.11
to the right-hand side of the ODEs of Eq. (41) over the sampling 14.18 14.08 14.07
period ( = 0.01 h) and the bounds are given for each element of 14.17 14.06 0.36*
the noise vector (|wi | ≤ i for i = 1, 2). Through extensive closed- 14.19 14.06 14.10
loop simulations of the CSTR under the controller k(x) and under
the LEMPC (described below) and with many realizations of the
6.1. RTO and EMPC
process noise, the set ˝e , a set where time-varying operation is
allowed while boundedness in ˝ is maintained, was determined
Throughout the EMPC literature, EMPC has been widely
to be e = 59.325.
reported as a method that merges process economic optimization
The first differential equation of Eq. (41) (CA ) is input-to-
and process control. Indeed, it does have this property. However,
state-stable (ISS) with respect to T. Therefore, a contractive
among all the current theoretical work on EMPC, only the work
Lyapunov-based constraint can be applied to the LEMPC to ensure
of Grüne [64] does not employ the use of a precomputed steady-
that the temperature converges to a neighborhood of the opti-
state or periodic operating trajectory as a terminal constraint or
mal steady-state temperature value. Namely, we define: VT (k ) :=
stability region. Notice that the EMPC schemes of Eq. (14) and (30),
(T (k ) − Ts∗ )2 . The LEMPC formulation is the same as Eq. (47) (mod-
each with provable stability properties, use precomputed infor-
ified as noted above to account for noise) with the following added
mation as constraints in the EMPC formulation. This information
constraints:
must come from some higher level information technology sys-
T (t) ≤ Tmax (50a) tem. Moreover, a wide variety of assumed a priori knowledge exists
amongst the various EMPC schemes. Some EMPC schemes only
V (x̃(t)) ≤ e ∀ t ∈ [0, N ) (50b) require the economically optimal steady-state (e.g., EMPC with
a terminal constraint), while others require much more knowl-
∂VT (k ) ∂VT (k ) edge like the optimal cyclical (perhaps periodic) operating strategy
f2 (x̃(0), u(0), 0) ≤ f2 (x̃(0), k(x̃(0)), 0) (50c)
∂T ∂T [79] and sufficient understanding of the behavior of the future
where f2 (·) is the right-hand side of the second ODE of Eq. (47). The economic pricing to model it appropriately [130]. Given the avail-
CSTR was initialized at many states distributed throughout state- ability of inexpensive computation, EMPC should take advantage
space including some cases where the initial state is outside ˝u . of any available information (like information provided by RTO)
The LEMPC described above was applied to the CSTR with an oper- to improve the closed-loop performance. Furthermore, EMPC does
ating period over which to enforce the average input constraint of not replace all the tasks completed by the RTO layer (e.g., data vali-
M = 20 and a prediction horizon of N = 20. Several simulations of dation and reconciliation and model updating). Thus, it is important
50.0 h length of operation were completed. In all cases, the LEMPC to keep in mind that EMPC does not entirely replace RTO.
was able to force the system to ˝ and maintain operation inside Remark 5. While EMPC may not completely replace RTO, it is
˝ without violating the state constraint. The closed-loop state important to understand that this does not imply that one should
trajectories over the first 1.0 h are shown in Fig. 14 for one initial not apply EMPC. Using EMPC will yield different closed-loop oper-
condition starting inside ˝ and one starting outside ˝u . More- ating trajectories compared to using (frequent) RTO (i.e., RTO that
over, the CSTR was simulated with the same realization of the is executed more frequently than that of traditional RTO systems)
process noise and same initial condition under the controller k(x) with a traditional control structure (e.g., tracking MPC). This is
applied in a sample-and-hold fashion and under a constant input because the RTO layer typically uses a steady-state process model,
equal to the steady-state input. The average economic cost over while the EMPC uses a dynamic process model. If Dynamic RTO (D-
each of these simulations is reported in Table 4. From these results, RTO) is used (i.e., RTO with a dynamic process model), it would be
an average of 0.6% closed-loop performance benefit was observed expected that the closed-loop economic performance under EMPC
with the LEMPC over the controller k(x) and the constant input u∗s . is better than that with D-RTO since D-RTO is typically executed at a
It is important to note that for one of the simulations that was ini- slower frequency than EMPC. If D-RTO is executed at the same rate
tialized outside ˝u the CSTR under the constant input u∗s settled as EMPC, then D-RTO, in this case, is essentially (one-layer) EMPC
on an offsetting steady-state which is denoted with an asterisk in (see Section 7.2 for more discussion on this point).
Table 4.
6.2. Closed-loop operation under EMPC
6. Discussion of current status of EMPC
From our experiences applying EMPC to various applications,
In this section, we reflect on the current status of EMPC devel- we have observed that the most closed-loop performance benefit
opments as well as unify some of the results on EMPC to compare under EMPC occurs when EMPC dictates a time-varying operating
and contrast the various approaches. policy which may range from a periodic or cyclical operating
1174 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178

policy to more complex non-periodic time-varying operation. was extended to the case of output feedback based on a high-gain
When steady-state operation is the optimal operating strategy, observer. While using high-gain observers may allow for proving
one should carefully consider the applicability of EMPC. Specifi- closed-loop stability, it may also be sensitive to measurement noise.
cally, when operating conditions or economic factors are updated As another approach to output feedback EMPC, moving horizon
infrequently and a system can be maintained near the optimal estimation (MHE) based on least squares techniques has become
steady-state (i.e., a sufficient time-scale separation exists between a popular state estimation technique because of its ability to
the frequency of economic factors update and the time constants handle nonlinear systems, account for the presence of distur-
of the process dynamics), it is unexpected that much benefit will bances, and account for constraints on decision variables leading to
be observed under EMPC since the system spends little time in the improved estimation performance [124,138,139,78,144]. One par-
transient phase relative to the total length of operation. For large- ticular formulation of MHE, robust MHE (RMHE), has been proposed
scale systems optimally operated at steady-state, one must also in [101,175] which is based on an auxiliary nonlinear observer
consider the computational burden required to solve EMPC on-line that asymptotically tracks the nominal system state. The auxiliary
which may be significant. For instance, the application of EMPC deterministic nonlinear observer is taken advantage of to calculate
to a large-scale chemical process network used in the production a confidence region that contains the actual system state taking into
of vinyl acetate led to comparable closed-loop performance when account bounded model uncertainties at every sampling time. The
compared to a well-tuned MPC formulated with a quadratic cost, region is then used to design a constraint on the state estimate in the
as economically optimal steady-state operation was likely optimal RMHE. The RMHE brings together deterministic and optimization-
[162]. However, the computation requirement of EMPC compared based observer design techniques. It was proved to give bounded
to MPC was considerably higher. If significant disturbances are estimation error in the case of bounded model uncertainties. In [55],
present such that it is difficult to achieve operation near the steady- an RMHE-based output feedback LEMPC was presented and stabil-
state (and steady-state operation is the optimal operating strategy), ity in the presence of measurement and process noise was proved.
one may consider using a local approximation of the nonlinear Future work in this direction should embark on considering the
model (e.g., linear dynamic model) and a local approximation of rigorous design of other types of output feedback EMPC schemes.
the economic cost when these approximations provide sufficient
accuracy of the process dynamics and economic cost, respectively. 7.2. Distributed EMPC, hierarchical EMPC, and distributed
This may convert the optimization to an easier problem (i.e., milder economic optimization
nonlinearities, possibly convert the optimization problem of EMPC
to a convex optimization problem which is readily solvable, etc.) In general, it has been pointed out that the computational bur-
helping to ease the computational burden of EMPC, while achieving den of EMPC over conventional MPC may be significantly higher
some possible closed-loop performance benefit. since EMPC may use a general nonlinear, non-convex cost function
with a sufficiently long prediction horizon for good closed-loop per-
6.3. Economic assessment of EMPC formance (e.g., [162]). In fact, the computational time required to
solve the EMPC may be greater than the time available (i.e., greater
For the chemical process example presented in this work, a than the sampling period). Furthermore, it has been argued in [166]
2% greater production rate was achieved under EMPC with nom- that the use of a one-layer EMPC which fully combines the RTO and
inal operation than with steady-state operation, while only a 0.6% supervisory (MPC) layers in Fig. 1 is undesirable within the context
greater production rate was achieved when process noise affected of industrial application without the use of some additional safety
the evolution of the process system. Whether it is an economi- control layer. Three potentially attractive choices to handle com-
cally viable option to apply EMPC to this particular process largely putational concerns include using distributed EMPC, hierarchical
depends on how valuable the product is and what the production EMPC, distributed optimization techniques, and/or any combina-
rate currently being realized in practice under the current control tion of these three approaches.
methodology is. Therefore, a careful economic assessment would It is clear in the context of MPC of large-scale process networks
need to be completed when determining if the benefit of apply- that distributed MPC (DMPC) schemes may significantly reduce the
ing EMPC is worth the engineering, capital, and related investment on-line computational load of MPC and thus, make MPC a feasible
costs (see the survey paper [14] for approaches for carrying out control methodology for large-scale, nonlinear process networks
such an evaluation). (e.g., see, for instance, [145,103,102,34] and the references therein).
While early work in this direction in the context of distributed
7. Future research directions EMPC (DEMPC) has shown promising results [29,42,95,96], more
work is necessary which includes work on the development of
In this section, we discuss some topics for future research work novel DEMPC algorithms, rigorous theoretical stability analysis, and
in the area of EMPC based on our experiences, observations, and introducing control loop decomposition methodologies for DEMPC.
motivations. The topics reflect our own bias and the list is certainly One potentially interesting research direction is to define new con-
not complete. trol loop decomposition methods on the basis of process economics.
This idea has some similarities to the so-called self-optimizing
7.1. State-estimation-based EMPC control methodology for control structure design for steady-state
operated processes [154,155,10,136]. However, it remains to be
Almost all of the proposed EMPC schemes rely on state feed- seen how these methods can be extended to dynamically operated
back. However, in practice, only measured output feedback may be systems.
available. Since there exists no separation principle for general non- Another alternative to using a single-layer EMPC system to com-
linear systems, it is hard to prove stability of the closed-loop system pute the control actions directly for the manipulated inputs, which
under a state-estimator or state-observer with a state feedback may be computationally taxing, is to use EMPC in a hierarchical
controller. An approach within nonlinear systems for the design control structure which is also commonly referred to as Dynamic-
of an output feedback controller is to use a high-gain observer with RTO [168,166]. Here, the idea is to maintain the current hierarchical
a stabilizing state feedback controller. In this case, one can apply control structure (Fig. 1), but replace the RTO layer with essentially
singular perturbation arguments to prove stability of the closed- an EMPC that computes an optimal operating trajectory. The opti-
loop system [86,87]. In a previous work [69], the LEMPC design mal trajectory is sent to the lower control layers to steer the process
 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178 1175

system to operate along these trajectories. Similar to current RTO, result in a computationally-intractable optimization problem to
the upper-layer EMPC would recompute its operating trajectory solve on-line. However, despite the demonstration of the computa-
for the system infrequently to lower computational requirements tional benefits of using reduced-order models in the formulation of
(i.e., not every sampling period), and thus, a loss of economic per- EMPC [91,90], rigorous theoretical stability analysis of PDEs under
formance may be expected over single-layer EMPC. Furthermore, EMPC remains an open topic.
stability analysis of the entire closed-loop hierarchical EMPC struc- Another class of systems that remains an open research topic
ture is in order although only limited work has been completed in within the context of EMPC is hybrid systems. Hybrid systems are
this direction (e.g., [50,51]). Future research in this direction should systems that are modeled with states that evolve on the continuous
include detailed stability analysis of the closed-loop structure and time-scale as well as states that evolve on a discrete time-scale like
novel algorithms that work to minimize the performance loss of discrete events. Historically, hybrid systems have attracted much
hierarchical EMPC compared to single-layer EMPC. attention within the control community (e.g., [17,32]). However,
Lastly, a continued desire within many fields that solve large- EMPC schemes for hybrid systems have received very limited atten-
scale, non-convex, nonlinear optimization problems is to continue tion. Within chemical process control, hybrid systems arise due to,
to push the boundaries of nonlinear optimization solver capabil- for instance, grade changes in the desired product (i.e., changes in
ities and computational efficiency. Thus, it will be important to product specifications), raw material changes, and variable energy
continue research efforts in parallel and distributed computation source pricing. Therefore, it is important to introduce EMPC meth-
(e.g., [18]) with a specific focus on distributed and parallel dynamic ods with guaranteed stability properties that may be applied to
optimization methods for EMPC. hybrid systems. Future research in the direction of distributed
parameter and hybrid systems may include proposing novel EMPC
7.3. Real-time calculation and network considerations schemes for these classes of systems and deriving conditions under
which stability and improved closed-loop performance of the sys-
In practice, an optimization-based controller takes a finite tem under EMPC may be guaranteed.
amount of time to solve which may be significant or insignifi-
cant depending on the time constants of the process dynamics.
Therefore, there is a (theoretical) maximum amount of time that
7.5. EMPC with input/output models
the nonlinear solver may spend in computation and must return a
control action by this maximum amount of time to ensure closed-
As its name implies, EMPC requires the availability of a dynamic
loop stability. Even if novel EMPC algorithms are presented that
model to compute its control actions. For the cases where the devel-
reduce the on-line computational load, it is still possible that the
opment of a sufficiently accurate first-principles dynamic model is
nonlinear optimization solver will not converge to a solution in
not possible, system identification techniques may need to be used
the time allotted. This may happen, for instance, if a poor initial
to obtain an accurate empirical input/output model of the process
guess is supplied to the solver. In this case, the solver will return
dynamics. In particular, nonlinear autoregressive moving average
a suboptimal input solution and thus, this type of MPC implemen-
with exogenous inputs (NARMAX) models may be one of many
tation is often referred to as suboptimal MPC (e.g., [146]). Future
types of nonlinear system identification techniques employed and
research should understand the stability properties of the input
used to construct an empirical model [20]. Future research should
solution computed by suboptimal EMPC. Furthermore, it may also
incorporate input/output models into the formulation of EMPC and
be desirable to look at methods that efficiently store previous input
investigate the capabilities and limitations of applying these mod-
solutions in a database as data storage is becoming increasingly
els in the context of EMPC.
inexpensive and take advantage of the database to provide the
solver with a potentially better initial guess.
Also of interest within the context of solving EMPC in real-time
is the fact that the components of the control architecture are con- 7.6. EMPC and safety/robustness considerations
nected through wired and/or wireless communication connections.
Communication delays between components may occur. Further- Operating in a continuously dynamic fashion, as EMPC may dic-
more, asynchronous measurements (i.e., asynchronous sampling) tate, may have considerable safety implications both positive and
may occur in certain practical applications. For example, species negative. On the positive side, a dynamically operated system may
concentration may be asynchronously measured. Therefore, EMPC offer some insight into the health of the components. For instance,
schemes that explicitly account for network considerations and one common method for fault detection within steady-state opera-
asynchronous sampling are important practical challenges of tion is to excite the process (induce a transient phase) and observe
EMPC. Within the context of conventional MPC, several results have its response in an effort to detect and isolate faulty components.
been obtained in this direction (e.g., [33]). It will be important to Since a system under EMPC may be under a constant “excited” state,
leverage these results and extend these results to EMPC. future research effort may focus on harnessing this for fault detec-
tion, isolation, and control reconfiguration to handle various types
7.4. EMPC of distributed parameter and hybrid systems of process faults. A potential drawback within the context of safety
is that the operating policy dictated by the EMPC must be robust to
Almost all of the work on control of distributed parameter sys- component failures. For example, the EMPC should compute oper-
tems modeled by PDEs has focused on steady-state stabilization ating trajectories that are safe with respect to potential component
and operation [30] especially in the context of predictive control failures. Furthermore, EMPC and process monitoring tools should
formulated for PDE systems (e.g. [47,45,43,44]). To this end, only be developed and deployed to help assess the overall process safety.
recently has some work been done on applying EMPC to PDE sys- While a few EMPC formulations with provable stability prop-
tems [91,90] which has primarily focused on the construction of erties in the presence of disturbances have been proposed (e.g.,
reduced-order models for EMPC by applying Galerkin’s method [68,77,39,67]), more work in this direction is in order. Future work
using analytical or empirical eigenfunctions as basis functions. It on EMPC should strive to provide provable stability and perfor-
is important to note that using a high-order spatial discretization mance of EMPC in the presence of disturbances as well as present
of the PDE model to obtain a system of ODEs describing the tem- novel EMPC algorithms and formulations that account for distur-
poral evolution of the PDE system in an EMPC framework may bances of known form and process noise with known statistics.
1176 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178

Acknowledgments [32] P.D. Christofides, N.H. El-Farra, Control of Nonlinear and Hybrid Process Sys-
tems: Designs for Uncertainty, Constraints and Time-Delays, Springer-Verlag,
Berlin, Germany, 2005.
Financial support from the National Science Foundation and the [33] P.D. Christofides, J. Liu, D. Muñoz de la Peña, Networked and Distributed
Department of Energy is gratefully acknowledged. Predictive Control: Methods and Nonlinear Process Network Applications.
Advances in Industrial Control Series, Springer-Verlag, London, England,
2011.
References [34] P.D. Christofides, R. Scattolini, D. Muñoz de la Peña, J. Liu, Distributed model
predictive control: a tutorial review and future research directions, Comput.
[1] O. Adeodu, D.J. Chmielewski, Control of electric power transmission networks Chem. Eng. 51 (2013) 21–41.
with massive energy storage using economic MPC, in: Proceedings of the [35] M.L. Darby, M. Nikolaou, J. Jones, D. Nicholson, RTO: an overview and assess-
American Control Conference, Washington, DC, 2013, pp. 5839–5844. ment of current practice, J. Process Control 21 (2011) 874–884.
[2] V. Adetola, M. Guay, Integration of real-time optimization and model predic- [36] J. Davis, T. Edgar, J. Porter, J. Bernaden, M. Sarli, Smart manufacturing, man-
tive control, J. Process Control 20 (2010) 125–133. ufacturing intelligence and demand-dynamic performance, Comput. Chem.
[3] T. Alamo, A. Ferramosca, A.H. González, D. Limon, D. Odloak, A gradient-based Eng. 47 (2012) 145–156.
strategy for integrating real time optimizer (RTO) with model predictive con- [37] M.T. de Gouvêa, D. Odloak, One-layer real time optimization of LPG production
trol (MPC), in: Proceedings of the 4th IFAC Nonlinear Model Predictive Control in the FCC unit: procedure, advantages and disadvantages, Comput. Chem.
Conference, Leeuwenhorst, Netherlands, 2012, pp. 33–38. Eng. 22 (Suppl. 1) (1998) S191–S198.
[4] R. Amrit, J.B. Rawlings, D. Angeli, Economic optimization using model predic- [38] M. Diehl, R. Amrit, J.B. Rawlings, A Lyapunov function for economic optimizing
tive control with a terminal cost, Ann. Rev. Control 35 (2011) 178–186. model predictive control, IEEE Trans. Automat. Control 56 (2011) 703–707.
[5] R. Amrit, J.B. Rawlings, L.T. Biegler, Optimizing process economics online using [39] L. Dobos, A. Király, J. Abonyi, Economic-oriented stochastic optimization in
model predictive control, Comput. Chem. Eng. 58 (2013) 334–343. advanced process control of chemical processes, Sci. World J. 2012 (2012), 10
[6] D. Angeli, R. Amrit, J.B. Rawlings, On average performance and stability of pp.
economic model predictive control, IEEE Trans. Automat. Control 57 (2012) [40] R. Dorfman, P. Samuelson, R. Solow, Linear Programming and Economic Anal-
1615–1626. ysis, McGraw-Hill, New York, 1958.
[7] T. Backx, O. Bosgra, W. Marquardt, Integration of model predictive control and [41] J.M. Douglas, Periodic reactor operation, Ind. Eng. Chem. Process Des. Dev. 6
optimization of processes: enabling technology for market driven process (1967) 43–48.
operation, in: Proceedings of the IFAC Symposium on Advanced Control of [42] P.A.A. Driessen, R.M. Hermans, P.P.J. van den Bosch, Distributed economic
Chemical Processes, Pisa, Italy, 2000, pp. 249–260. model predictive control of networks in competitive environments, in:
[8] J.E. Bailey, Periodic operation of chemical reactors: a review, Chem. Eng. Com- Proceedings of the 51st IEEE Conference on Decision and Control, Maui, HI,
mun. 1 (1973) 111–124. 2012, pp. 266–271.
[9] J.E. Bailey, F.J.M. Horn, Comparison between two sufficient conditions for [43] S. Dubljevic, P.D. Christofides, Predictive control of parabolic PDEs with
improvement of an optimal steady-state process by periodic operation, J. boundary control actuation, Chem. Eng. Sci. 61 (2006) 6239–6248.
Optim. Theory Appl. 7 (1971) 378–384. [44] S. Dubljevic, P.D. Christofides, Predictive output feedback control of parabolic
[10] M. Baldea, A. Araujo, S. Skogestad, P. Daoutidis, Dynamic considerations partial differential equations (PDEs), Ind. Eng. Chem. Res. 45 (2006)
in the synthesis of self-optimizing control structures, AIChE J. 54 (2008) 8421–8429.
1830–1841. [45] S. Dubljevic, N.H. El-Farra, P. Mhaskar, P.D. Christofides, Predictive control
[11] M. Baldea, P. Daoutidis, Control of integrated process networks - A multi-time of parabolic PDEs with state and control constraints, Int. J. Robust Nonlinear
scale perspective, Comput. Chem. Eng. 31 (2007) 426–444. Control 16 (2006) 749–772.
[12] M. Baldea, C.R. Touretzky, Nonlinear model predictive control of [46] S. Dubljević, N. Kazantzis, A new Lyapunov design approach for nonlinear
energy-integrated process systems, Syst. Control Lett. 62 (2013) systems based on Zubov’s method, Automatica 38 (2002) 1999–2007.
723–731. [47] S. Dubljevic, P. Mhaskar, N.H. El-Farra, P.D. Christofides, Predictive con-
[13] Y. Barshad, E. Gulari, A dynamic study of CO oxidation on supported platinum, trol of transport-reaction processes, Comput. Chem. Eng. 29 (2005)
AIChE J. 31 (1985) 649–658. 2335–2345.
[14] M. Bauer, I.K. Craig, Economic assessment of advanced process control - A [48] N.H. El-Farra, P.D. Christofides, Bounded robust control of constrained multi-
survey and framework, J. Process Control 18 (2008) 2–18. variable nonlinear processes, Chem. Eng. Sci. 58 (2003) 3025–3047.
[15] V.M. Becerra, P.D. Roberts, G.W. Griffiths, Novel developments in pro- [49] M. Ellis, P.D. Christofides, Economic model predictive control with time-
cess optimisation using predictive control, J. Process Control 8 (1998) varying objective function for nonlinear process systems, AIChE J. 60 (2014)
117–138. 507–519.
[16] R.E. Bellman, Dynamic Programming, Princeton University Press, Princeton, [50] M. Ellis, P.D. Christofides, Integrating dynamic economic optimization and
NJ, 1957. model predictive control for optimal operation of nonlinear process systems,
[17] A. Bemporad, M. Morari, Control of systems integrating logic, dynamics, and Control Eng. Pract. 22 (2014) 242–251.
constraints, Automatica 35 (1999) 407–427. [51] M. Ellis, P.D. Christofides, Optimal time-varying operation of nonlinear pro-
[18] D.P. Bertsekas, J.N. Tsitsiklis, Parallel and Distributed Computation: Numerical cess systems with economic model predictive control, Ind. Eng. Chem. Res.
Methods, Athena Scientific, Belmont, MA, 1997. 53 (2014) 4991–5001.
[19] L.T. Biegler, Nonlinear Programming: Concepts, Algorithms, and Applications [52] M. Ellis, P.D. Christofides, Performance monitoring of economic
to Chemical Processes, SIAM, Philadelphia, PA, 2010. model predictive control systems. Ind. Eng. Chem. Res. (in press),
[20] S.A. Billings, Nonlinear System Identification: NARMAX Methods in the Time, http://dx.doi.org/10.1021/ie403462y
Frequency, and Spatio-Temporal Domains, John Wiley & Sons, 2013. [53] M. Ellis, P.D. Christofides, On finite-time and infinite-time cost improvement
[21] S. Bittanti, G. Fronza, G. Guardabassi, Periodic control: a frequency domain of economic model predictive control for nonlinear systems. Automatica (sub-
approach, IEEE Trans. Automat. Control 18 (1973) 33–38. mitted for publication).
[22] C. Brosilow, G.Q. Zhao, A linear programming approach to constrained mul- [54] M. Ellis, M. Heidarinejad, P.D. Christofides, Economic model predictive con-
tivariable process control, in: C.T. Leondes (Ed.), System Identification and trol of nonlinear singularly perturbed systems, J. Process Control 23 (2013)
Adaptive Control, Part 3 of 3, volume 27 of Control and Dynamic Systems, 743–754.
Academic Press, 1988, pp. 141–181. [55] M. Ellis, J. Zhang, J. Liu, P.D. Christofides, Robust moving horizon estimation
[23] H. Budman, M. Kzyonsek, P. Silveston, Control of a nonadiabatic packed based output feedback economic model predictive control. Syst. Control Lett.
bed reactor under periodic flow reversal, Can. J. Chem. Eng. 74 (1996) (in press), http://dx.doi.org/10.1016/j.sysconle.2014.03.003
751–759. [56] S. Engell, Feedback control for optimal process operation, J. Process Control
[24] H. Budman, P.L. Silveston, Control of periodically operated reactors, Chem. 17 (2007) 203–219.
Eng. Sci. 63 (2008) 4942–4954. [57] L. Fagiano, A.R. Teel, Generalized terminal state constraint for model predic-
[25] C.I. Byrnes, W. Lin, Losslessness, feedback equivalence, and the global stabi- tive control, Automatica 49 (2013) 2622–2631.
lization of discrete-time nonlinear systems, IEEE Trans. Automat. Control 39 [58] A. Ferramosca, J.B. Rawlings, D. Limon, E.F. Camacho, Economic MPC for a
(1994) 83–98. changing economic criterion, in: Proceedings of the 49th IEEE Conference on
[26] E.F. Camacho, C.B. Alba, Model predictive control, 2nd ed., Springer, 2013. Decision and Control, Alanta, GA, 2010, pp. 6131–6136.
[27] D.A. Carlson, A.B. Haurie, A. Leizarowitz, Infinite Horizon Optimal Control, [59] W. Findeisen, F.N. Bailey, M. Brdyś, K. Malinowski, P. Tatjewski, A. Woźniak,
Springer Verlag, 1991. Control and Coordination in Hierarchical Systems, John Wiley & Sons, New
[28] D. Cass, Optimum growth in an aggregative model of capital accumulation: a York, 1980.
turnpike theorem, Econometrica 34 (1966) 833–850. [60] J.F. Forbes, T.E. Marlin, Model accuracy for economic optimizing controllers:
[29] X. Chen, M. Heidarinejad, J. Liu, P.D. Christofides, Distributed economic MPC: the bias update case, Ind. Eng. Chem. Res. 33 (1994) 1919–1929.
application to a nonlinear chemical process network, J. Process Control 22 [61] J.F. Forbes, T.E. Marlin, J.F. MacGregor, Model adequacy requirements for opti-
(2012) 689–699. mizing plant operations, Comput. Chem. Eng. 18 (1994) 497–510.
[30] P.D. Christofides, Nonlinear and Robust Control of PDE Systems: Methods and [62] C.E. García, D.M. Prett, M. Morari, Model predictive control: theory and
Applications to Transport-Reaction Processes, Birkhäuser, Boston, 2001. practice – a survey, Automatica 25 (1989) 335–348.
[31] P.D. Christofides, J.F. Davis, N.H. El-Farra, D. Clark, K.R.D. Harris, J.N. Gipson, [63] A. Gopalakrishnan, L.T. Biegler, Economic nonlinear model predictive control
Smart plant operations: vision, progress and challenges, AIChE J. 53 (2007) for periodic optimal operation of gas pipeline networks, Comput. Chem. Eng.
2734–2741. 52 (2013) 90–99.
 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178 1177

[64] L. Grüne, Economic receding horizon control without terminal constraints, [92] L. Lao, M. Ellis, P.D. Christofides, Smart Manufacturing: handling preventive
Automatica 49 (2013) 725–734. actuator maintenance and economics using model predictive control. AIChE
[65] L. Grüne, J. Pannek, K. Worthmann, D. Nešić, Redesign techniques for J. (in press), http://dx.doi.org/10.1002/aic.14427
nonlinear sampled-data systems, at-Automatisierungstechnik 56 (2008) [93] M. Ławryńczuk, P.M. Marusak, P. Tatjewski, Cooperation of model predictive
38–48. control with steady-state economic optimisation, Control Cybern. 37 (2008)
[66] G. Guardabassi, A. Locatelli, S. Rinaldi, Status of periodic optimization of 133–158.
dynamical systems, J. Optim. Theory Appl. 14 (1974) 1–20. [94] C.K. Lee, J.E. Bailey, Modification of consecutive-competitive reaction selec-
[67] M. Guay, V. Adetola, Adaptive economic optimising model predic- tivity by periodic operation, Ind. Eng. Chem. Process Des. Dev. 19 (1980)
tive control of uncertain nonlinear systems, Int. J. Control 86 (2013) 160–166.
1425–1437. [95] J. Lee, D. Angeli, Cooperative distributed model predictive control for linear
[68] M. Heidarinejad, J. Liu, P.D. Christofides, Economic model predictive control plants subject to convex economic objectives, in: Proceedings of the 50th
of nonlinear process systems using Lyapunov techniques, AIChE J. 58 (2012) IEEE Conference on Decision and Control and European Control Conference,
855–870. Orlando, FL, 2011, pp. 3434–3439.
[69] M. Heidarinejad, J. Liu, P.D. Christofides, State-estimation-based economic [96] J. Lee, D. Angeli, Distributed cooperative nonlinear economic MPC, in:
model predictive control of nonlinear systems, Syst. Control Lett. 61 (2012) Proceedings of the 20th International Symposium on Mathematical Theory
926–935. of Networks and Systems, Melbourne, Australia, 2012.
[70] M. Heidarinejad, J. Liu, P.D. Christofides, Algorithms for improved fixed-time [97] J.H. Lee, S. Natarajan, K.S. Lee, A model-based predictive control approach to
performance of Lyapunov-based economic model predictive control of non- repetitive control of continuous processes with periodic operations, J. Process
linear systems, J. Process Control 23 (2013) 404–414. Control 11 (2001) 195–207.
[71] M. Heidarinejad, J. Liu, P.D. Christofides, Economic model predictive control [98] R.M. Lima, I.E. Grossmann, Y. Jiao, Long-term scheduling of a single-unit multi-
of switched nonlinear systems, Syst. Control Lett. 62 (2013) 77–84. product continuous process to manufacture high performance glass, Comput.
[72] A. Helbig, O. Abel, W. Marquardt, Structural concepts for optimization based Chem. Eng. 35 (2011) 554–574.
control of transient processes, in: F. Allgöwer, A. Zheng (Eds.), Nonlinear [99] Y. Lin, E. Sontag, Y. Wang, A smooth converse Lyapunov theorem for robust
Model Predictive Control, volume 26 of Progress in Systems and Control stability, SIAM J. Control Optim. 34 (1996) 124–160.
Theory, Birkhäuser, Basel, 2000, pp. 295–311. [100] Y. Lin, E.D. Sontag, A universal formula for stabilization with bounded controls,
[73] T.G. Hovgaard, S. Boyd, L.F.S. Larsen, J.B. Jørgensen, Nonconvex model Syst. Control Lett. 16 (1991) 393–397.
predictive control for commercial refrigeration, Int. J. Control 86 (2013) [101] J. Liu, Moving horizon state estimation for nonlinear systems with bounded
1349–1366. uncertainties, Chem. Eng. Sci. 93 (2013) 376–386.
[74] T.G. Hovgaard, K. Edlund, J.B. Jørgensen, The potential of economic MPC for [102] J. Liu, X. Chen, D. Muñoz de la Peña, P.D. Christofides, Sequential and iterative
power management, in: Proceedings of the 49th IEEE Conference on Decision architectures for distributed model predictive control of nonlinear process
and Control, Altanta, GA, 2010, pp. 7533–7538. systems, AIChE J. 56 (2010) 2137–2149.
[75] T.G. Hovgaard, L.F.S. Larsen, K. Edlund, J.B. Jørgensen, Model predictive control [103] J. Liu, D. Muñoz de la Peña, P.D. Christofides, Distributed model predictive
technologies for efficient and flexible power consumption in refrigeration control of nonlinear process systems, AIChE J. 55 (2009) 1171–1184.
systems, Energy 44 (2012) 105–116. [104] J. Ma, J. Qin, T. Salsbury, P. Xu, Demand reduction in building energy sys-
[76] T.G. Hovgaard, L.F.S. Larsen, J.B. Jørgensen, Robust economic MPC for a tems based on economic model predictive control, Chem. Eng. Sci. 67 (2012)
power management scenario with uncertainties, in: Proceedings of the 50th 92–100.
IEEE Conference on Decision and Control and European Control Conference, [105] E. Mancusi, P. Altimari, L. Russo, S. Crescitelli, Multiplicities of temperature
Orlando, FL, 2011, pp. 1515–1520. wave trains in periodically forced networks of catalytic reactors for reversible
[77] R. Huang, L.T. Biegler, E. Harinath, Robust stability of economically oriented exothermic reactions, Chem. Eng. J. 171 (2011) 655–668.
infinite horizon NMPC that include cyclic processes, J. Process Control 22 [106] T.E. Marlin, A.N. Hrymak, Real-time operations optimization of continuous
(2012) 51–59. processes, in: Proceedings of the Fifth International Conference on Chemical
[78] R. Huang, L.T. Biegler, S.C. Patwardhan, Fast offset-free nonlinear model pre- Process Control, Tahoe City, CA, 1996, pp. 156–164.
dictive control based on moving horizon estimation, Ind. Eng. Chem. Res. 49 [107] W. Marquardt, Nonlinear model reduction for optimization based control
(2010) 7882–7890. of transient chemical processes, in: Proceedings of the Sixth Inter-
[79] R. Huang, E. Harinath, L.T. Biegler, Lyapunov stability of economically oriented national Conference on Chemical Process Control, Tucson, AZ, 2001,
NMPC for cyclic processes, J. Process Control 21 (2011) 501–509. pp. 12–42.
[80] E.A.N. Idris, S. Engell, Economics-based NMPC strategies for the operation [108] J.L. Massera, Contributions to stability theory, Ann. Math. 64 (1956)
and control of a continuous catalytic distillation process, J. Process Control 22 182–206.
(2012) 1832–1843. [109] D.Q. Mayne, J.B. Rawlings, C.V. Rao, P.O.M. Scokaert, Constrained model
[81] J.V. Kadam, W. Marquardt, Integration of economical optimization and control predictive control: stability and optimality, Automatica 36 (2000)
for intentionally transient process operation, in: R. Findeisen, F. Allgöwer, L.T. 789–814.
Biegler (Eds.), Assessment and Future Directions of Nonlinear Model Predic- [110] L.W. McKenzie, Turnpike theory, Econometrica 44 (1976) 841–865.
tive Control, volume 358 of Lecture Notes in Control and Information Sciences, [111] L.W. McKenzie, Optimal economic growth, turnpike theorems and com-
Springer, Berlin/Heidelberg, 2007, pp. 419–434. parative dynamics, in: K.J. Arrow, M.D. Intriligator (Eds.), Handbook of
[82] J.V. Kadam, W. Marquardt, M. Schlegel, T. Backx, O.H. Bosgra, P.-J. Brouwer, Mathematical Economics, volume 3, Elsevier, 1986, pp. 1281–1355 (Chapter
G. Dünnebier, D. van Hessem, A. Tiagounov, S. de Wolf, Towards inte- 26).
grated dynamic real-time optimization and control of industrial processes, [112] D.I. Mendoza-Serrano, D.J. Chmielewski, HVAC control using infinite-horizon
in: Proceedings of the 4th International Conference of the Founda- economic MPC, in: Proceedings of the 51st IEEE Conference on Decision and
tions of Computer-Aided Process Operations, Coral Springs, FL, 2003, Control, Maui, Hawaii, 2012, pp. 6963–6968.
pp. 593–596. [113] D.I. Mendoza-Serrano, D.J. Chmielewski, Demand response for chemical man-
[83] J.V. Kadam, M. Schlegel, W. Marquardt, R.L. Tousain, D.H. van Hessem, J. van ufacturing using economic MPC, in: Proceedings of the American Control
den Berg, O.H. Bosgra, A two-level strategy of integrated dynamic optimiza- Conference, Washington, D.C., 2013, pp. 6655–6660.
tion and control of industrial processes – a case study, in: J. Grievink, J. van [114] P. Mhaskar, N.H. El-Farra, P.D. Christofides, Predictive control of switched
Schijndel (Eds.), Proceedings of the European Symposium on Computer Aided nonlinear systems with scheduled mode transitions, IEEE Trans. Automat.
Process Engineering-12, volume 10 of Computer Aided Chemical Engineering, Control 50 (2005) 1670–1680.
The Hague, The Netherlands, 2002, pp. 511–516. [115] P. Mhaskar, N.H. El-Farra, P.D. Christofides, Stabilization of nonlinear systems
[84] R.E. Kalman, Contributions to the theory of optimal control, Bol. Soc. Mat. with state and control constraints using Lyapunov-based predictive control,
Mex. 5 (1960) 102–119. Syst. Control Lett. 55 (2006) 650–659.
[85] I. Karafyllis, C. Kravaris, Global stability results for systems under sampled- [116] M. Morari, J.H. Lee, Model predictive control: past, present and future, Com-
data control, Int. J. Robust Nonlinear Control 19 (2009) 1105–1128. put. Chem. Eng. 23 (1999) 667–682.
[86] H.K. Khalil, High-gain observers in nonlinear feedback control, in: H. Nijmei- [117] A.M. Morshedi, C.R. Cutler, T.A. Skrovanek, Optimal solution of dynamic
jer, T.I. Fossen (Eds.), New Directions in nonlinear observer design, volume matrix control with linear programing techniques (LDMC), in: Proceedings
244 of Lecture Notes in Control and Information Sciences, Springer, London, of the American Control Conference, 1985, pp. 199–208.
1999, pp. 249–268. [118] D. Muñoz de la Peña, P.D. Christofides, Lyapunov-based model predictive con-
[87] H.K. Khalil, Nonlinear Systems, 3rd ed., Prentice Hall, Upper Saddle River, NJ, trol of nonlinear systems subject to data losses, IEEE Trans. Automat. Control
2002. 53 (2008) 2076–2089.
[88] P. Kokotović, M. Arcak, Constructive nonlinear control: a historical perspec- [119] M.A. Müller, F. Allgöwer, Robustness of steady-state optimality in economic
tive, Automatica 37 (2001) 637–662. model predictive control, in: Proceedings of the 51st IEEE Conference on
[89] P. Kunkel, O. von dem Hagen, Numerical solution of infinite-horizon optimal- Decision and Control, Maui, Hawaii, 2012, pp. 1011–1016.
control problems, Computat. Econ. 16 (2000) 189–205. [120] M.A. Müller, D. Angeli, F. Allgöwer, Economic model predictive control with
[90] L. Lao, M. Ellis, P.D. Christofides, Economic model predictive control of self-tuning terminal cost, Eur. J. Control 19 (2013) 408–416.
parabolic PDE systems: addressing state estimation and computational effi- [121] M.A. Müller, D. Angeli, F. Allgöwer, Economic model predictive control with
ciency, J. Process Control 24 (2014) 448–462. transient average constraints, in: Proceedings of the 52nd IEEE Conference on
[91] L. Lao, M. Ellis, P.D. Christofides, Economic model predictive con- Decision and Control, Florence, Italy, 2013, pp. 5119–5124.
trol of transport-reaction processes. Ind. Eng. Chem. Res. (in press), [122] M.A. Müller, D. Angeli, F. Allgöwer, On convergence of averagely con-
http://dx.doi.org/10.1021/ie401016a strained economic MPC and necessity of dissipativity for optimal steady-state
1178 M. Ellis et al. / Journal of Process Control 24 (2014) 1156–1178

operation, in: Proceedings of the American Control Conference, Washington, [149] X. Shu, K. Rigopoulos, A. Çinar, Vibrational control of an exothermic CSTR: pro-
D.C., 2013, pp. 3147–3152. ductivity improvement by multiple input oscillations, IEEE Trans. Automat.
[123] M.A. Müller, D. Angeli, F. Allgöwer, R. Amirt, J.B. Rawlings, Convergence in Control 34 (1989) 193–196.
economic model predictive control with average constraints. Automatica [150] J.J. Siirola, T.F. Edgar, Process energy systems: control, economic, and sus-
(submitted for publication). tainability objectives, Comput. Chem. Eng. 47 (2012) 134–144.
[124] K.R. Muske, J.B. Rawlings, J.H. Lee, Receding horizon recursive state estima- [151] P.L. Silveston, Periodic operation of chemical reactors - A review of the exper-
tion, in: Proceedings of the American Control Conference, San Francisco, CA, imental literature, Sādhanā 10 (1987) 217–246.
1993, pp. 900–904. [152] P.L. Silveston, R.R. Hudgins (Eds.), Periodic Operation of Reactors, Elsevier,
[125] K.R. Muske, Steady-state target optimization in linear model predictive Oxford, England, 2013.
control, in: Proceedings of the American Control Conference, volume 6, Albu- [153] P.L. Silveston, R.R. Hudgins, A. Renken, Periodic operation of catalytic reactors
querque, NM, 1997, pp. 3597–3601. - Introduction and overview, Catal. Today 25 (1995) 91–112.
[126] S. Natarajan, J.H. Lee, Repetitive model predictive control applied to a sim- [154] S. Skogestad, Plantwide control: the search for the self-optimizing control
ulated moving bed chromatography system, Comput. Chem. Eng. 24 (2000) structure, J. Process Control 10 (2000) 487–507.
1127–1133. [155] S. Skogestad, Self-optimizing control: the missing link between steady-state
[127] R. Nath, Z. Alzein, On-line dynamic optimization of olefins plants, Comput. optimization and control, Comput. Chem. Eng. 24 (2000) 569–575.
Chem. Eng. 24 (2000) 533–538. [156] E.D. Sontag, Mathematical Control Theory: Deterministic Finite Dimensional
[128] D. Nešić, A.R. Teel, Sampled-data control of nonlinear systems: an overview Systems, volume 6, Springer, 1998.
of recent results, in: S.O.R. Moheimani (Ed.), Perspectives in Robust Control, [157] G. De Souza, D. Odloak, A.C. Zanin, Real time optimization (RTO) with model
volume 268 of Lecture Notes in Control and Information Sciences, Springer, predictive control (MPC), Comput. Chem. Eng. 34 (2010) 1999–2006.
London, 2001, pp. 221–239. [158] L.E. Sterman, B.E. Ydstie, The steady-state process with periodic perturba-
[129] D. Nešić, A.R. Teel, D. Carnevale, Explicit computation of the sampling period tions, Chem. Eng. Sci. 45 (1990) 721–736.
in emulation of controllers for nonlinear sampled-data systems, IEEE Trans. [159] L.E. Sterman, B.E. Ydstie, Periodic forcing of the CSTR: an application of the
Automat. Control 54 (2009) 619–624. generalized -criterion, AIChE J. 37 (1991) 986–996.
[130] B.P. Omell, D.J. Chmielewski, IGCC power plant dispatch using infinite- [160] P. Tatjewski, Advanced control and on-line process optimization in multilayer
horizon economic model predictive control, Ind. Eng. Chem. Res. 52 (2013) structures, Ann. Rev. Control 32 (2008) 71–85.
3151–3164. [161] T. Tosukhowong, J.M. Lee, J.H. Lee, J. Lu, An introduction to a dynamic plant-
[131] F. Özgülşen, R.A. Adomaitis, A. Çinar, A numerical method for determining wide optimization strategy for an integrated plant, Comput. Chem. Eng. 29
optimal parameter values in forced periodic operation, Chem. Eng. Sci. 47 (2004) 199–208.
(1992) 605–613. [162] T.S. Tu, M. Ellis, P.D. Christofides, Model predictive control of a nonlinear large-
[132] F. Özgülşen, A. Çinar, Forced periodic operation of tubular reactors, Chem. scale process network used in the production of vinyl acetate, Ind. Eng. Chem.
Eng. Sci. 49 (1994) 3409–3419. Res. 52 (2013) 12463–12481.
[133] F. Özgülşen, S.J. Kendra, A. Çinar, Nonlinear predictive control of periodically [163] A. Wächter, L.T. Biegler, On the implementation of an interior-point filter line-
forced chemical reactors, AIChE J. 39 (1993) 589–598. search algorithm for large-scale nonlinear programming, Math. Program. 106
[134] A. Papachristodoulou, S. Prajna, On the construction of Lyapunov functions (2006) 25–57.
using the sum of squares decomposition, in: Proceedings of the 41st IEEE [164] A. Walther, Computing sparse hessians with automatic differentiation, ACM
Conference on Decision and Control, Las Vegas, NV, 2002, pp. 3482–3487. Trans. Math. Softw. 34 (2008), 15 pp.
[135] L.S. Pontryagin, V.G. Boltyanskii, R.V. Gamkrelidze, E.F. Mishchenko, Mathe- [165] J.C. Willems, Dissipative dynamical systems part I: General theory, Arch.
matical theory of optimal processes, Fizmatgiz, Moscow, 1961. Rational Mech. Anal. 45 (1972) 321–351.
[136] A. Psaltis, I.K. Kookos, C. Kravaris, Plant-wide control structure selec- [166] I.J. Wolf, D.A. Muñoz, W. Marquardt, Consistent hierarchical economic NMPC
tion methodology based on economics, Comput. Chem. Eng. 52 (2013) for a class of hybrid systems using neighboring-extremal updates, J. Process
240–248. Control (2013), http://dx.doi.org/10.1016/j.jprocont.2013.10.002.
[137] S.J. Qin, T.A. Badgwell, A survey of industrial model predictive control tech- [167] L. Würth, R. Hannemann, W. Marquardt, Neighboring-extremal updates for
nology, Control Eng. Pract. 11 (2003) 733–764. nonlinear model-predictive control and dynamic real-time optimization, J.
[138] C.V. Rao, J.B. Rawlings, J.H. Lee, Constrained linear state estimation-a moving Process Control 19 (2009) 1277–1288.
horizon approach, Automatica 37 (2001) 1619–1628. [168] L. Würth, R. Hannemann, W. Marquardt, A two-layer architecture for eco-
[139] C.V. Rao, J.B. Rawlings, D.Q. Mayne, Constrained state estimation for nonlinear nomically optimal process control and operation, J. Process Control 21 (2011)
discrete-time systems: stability and moving horizon approximations, IEEE 311–321.
Trans. Automat. Control 48 (2003) 246–258. [169] L. Würth, J.B. Rawlings, W. Marquardt, Economic dynamic real-time opti-
[140] J.B. Rawlings, Tutorial overview of model predictive control, IEEE Control Syst. mization and nonlinear model-predictive control on the infinite horizon, in:
Mag. 20 (2000) 38–52. Proceedings of the 7th IFAC International Symposium on Advanced Control
[141] J.B. Rawlings, R. Amrit, Optimizing process economic performance using of Chemical Processes, Istanbul, Turkey, 2007, pp. 219–224.
model predictive control, in: L. Magni, D.M. Raimondo, F. Allgöwer [170] L. Würth, I.J. Wolf, W. Marquardt, On the numerical solution of discounted
(Eds.), Nonlinear Model Predictive Control, volume 384 of Lecture Notes economic NMPC on infinite horizons, in: Proceedings of the 10th IFAC Inter-
in Control and Information Sciences, Springer Berlin Heidelberg, 2009, national Symposium on Dynamics and Control of Process Systems, Bombay,
pp. 119–138. Mumbai, India, 2013, pp. 209–214.
[142] J.B. Rawlings, D. Angeli, C.N. Bates, Fundamentals of economic model pre- [171] C.-M. Ying, B. Joseph, Performance and stability analysis of LP-MPC and QP-
dictive control, in: Proceedings of the 51st IEEE Conference on Decision and MPC cascade control systems, AIChE J. 45 (1999) 1521–1534.
Control, Maui, Hawaii, 2012, pp. 3851–3861. [172] C. Yousfi, R. Tournier, Steady state optimization inside model predictive
[143] J.B. Rawlings, D. Bonné, J.B. Jørgensen, A.N. Venkat, S.B. Jørgensen, Unreach- control, in: Proceedings of the American Control Conference, Boston, Mas-
able setpoints in model predictive control, IEEE Trans. Automat. Control 53 sachusetts, 1991, pp. 1866–1870.
(2008) 2209–2215. [173] A.C. Zanin, M.T. de Gouvêa, D. Odloak, Integrating real-time optimization
[144] J.B. Rawlings, L. Ji, Optimization-based state estimation: current status and into the model predictive controller of the FCC system, Control Eng. Pract.
some new results, J. Process Control 22 (2012) 1439–1444. 10 (2002) 819–831.
[145] R. Scattolini, Architectures for distributed and hierarchical model predictive [174] M. Zanon, S. Gros, M. Diehl, A Lyapunov function for periodic economic opti-
control - A review, J. Process Control 19 (2009) 723–731. mizing model predictive control, in: Proceedings of the 52nd IEEE Conference
[146] P.O.M. Scokaert, D.Q. Mayne, J.B. Rawlings, Suboptimal model predictive con- on Decision and Control, Florence, Italy, 2013, pp. 5107–5112.
trol (feasibility implies stability), IEEE Trans. Automat. Control 44 (1999) [175] J. Zhang, J. Liu, Lyapunov-based MPC with robust moving horizon estimation
648–654. and its triggered implementation, AIChE J. 59 (2013) 4273–4286.
[147] D.E. Seborg, T.F. Edgar, D.A. Mellichamp, F.J. Doyle, Process dynamics and [176] X. Zhu, W. Hong, S. Wang, Implementation of advanced control for a heat-
control, 3rd edition, Wiley, New York, NY, 2010. integrated distillation column system, in: Proceedings of the 30th Annual
[148] S.E. Sequeira, M. Graells, L. Puigjaner, Real-time evolution for on-line opti- Conference of IEEE Industrial Electronics Society, Busan, Korea, 2004, pp.
mization of continuous processes, Ind. Eng. Chem. Res. 41 (2002) 1815–1825. 2006–2011.